Add files using upload-large-folder tool

human_reviews/06ZvHHBR0i.md

@@ -0,0 +1,214 @@
+# Adversarial Multi-Agent Evaluation of Large Language Models through Iterative Debate
+
+- Decision: Reject
+- Scores: 3, 3, 1, 3
+
+## Abstract
+This paper explores optimal architectures for evaluating the outputs of large language models (LLMs) using LLMs themselves. We propose a novel framework that interprets LLMs as advocates within an ensemble of interacting agents, allowing them to defend their answers and reach conclusions through a judge and jury system. This approach offers a more dynamic and comprehensive evaluation process compared to traditional human-based assessments or automated metrics. We discuss the motivation behind this framework, its key components, and comparative advantages. We also present a probabilistic model to evaluate the error reduction achieved by iterative advocate systems. Finally, we outline experiments to validate the effectiveness of multi-advocate architectures and discuss future research directions.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+This paper presents two multi-agent systems inspired by courtroom for evaluating the outputs of LLMs. The experiments show the proposed frameworks improve accuracy compared with a single LLM as a judge.
+
+### Strengths
+The method incorporates insights from a legal decision-making perspective, and provide two frameworks that simulate the human workflow in court.
+
+### Weaknesses
+- Lack of experiments: More evaluation datasets and baselines should be incorporated into the experiments. For example, LLM-based multi-agent evaluators such as PRD [1] and ChatEval [2] could be baselines. There are many datasets in this LLM-based evaluator topic, such as AlignBench [3], AUTO-J [4] and LLMEval [5].
+
+- The presentation needs to be refined: 
+
+(a) The background (in both Section 1 and Section 2) is taking up too much space. This background can be concluded to make space for evaluation details in Appendix D. 
+
+**(b) The Conclusion section and Appendix A.5 are likely to be AI-generated (according to GPTZero).**
+
+- The multi-agent systems will surely use more tokens compared to LLM-as-a-judge. What is the cost per run compared to other multi-agent frameworks (such as PRD [1] and ChatEval [2])?
+
+[1] PRD: Peer Rank and Discussion Improve Large Language Model based Evaluations https://arxiv.org/abs/2307.02762
+
+[2] ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate https://arxiv.org/abs/2308.07201
+
+[3] AlignBench: Benchmarking Chinese Alignment of Large Language Models https://aclanthology.org/2024.acl-long.624/
+
+[4] Generative Judge for Evaluating Alignment https://arxiv.org/abs/2310.05470
+
+[5] LLMEval: A Preliminary Study on How to Evaluate Large Language Models https://ojs.aaai.org/index.php/AAAI/article/view/29934
+
+### Questions
+- The Qwen and Gemini model versions should be specified.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 2
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+5
+
+### Summary
+This work explores different processes of rating large language model (LLM) outputs using LLMs. Inspired by legal, psychological, and decision theory, the authors propose two such processes: (1) “Multi-Advocate One-Round Evaluations” (MORE) and “Single Advocate Multi-Round Evaluation” (“SAMRE”). Given a question and two (LLM) outputs, each process uses LLMs in different roles, e.g., as advocates, jurors, or judges, to (iteratively) select the “best” output. The authors further present two theorems respectively claiming that (1) aggregated multi-advocate arguments lead to greater score differentiation than those obtained using iterative debates, and (2) that multi-advocate argumentation requires fewer rounds of interaction to receive the same level of score differentiation as iterative debate. The two processes are tested on the MT-Bench dataset and compared to a baseline of a single LLM judge process using six different LLMs. The authors conclude that their experimental results provide strong empirical evidence for their proposed methods.
+
+### Strengths
+[clarity] The work was generally easy to read with crisp writing.  
+  
+[significance] Exploring ways to improve the evaluation of LLM outputs is an important research direction that was well motivated by the authors.
+
+### Weaknesses
+[originality] Several works have proposed different ways of using LLM ensembles to evaluate LLM outputs. While the authors spend considerable time discussing connections to various disciplines, e.g., decision theory, legal discourse, and psychology, few tangible insights are presented as to how this specific ensemble utilizes these disciplines.
+  
+[quality] The experimental results presented in this work simply do not pass the bar for this conference: (1) Only a single, limited dataset is used, (2) critical experimental details are missing, e.g., number of samples used, confidence intervals, temperatures, single-judge baseline setting, length of argument outputs, etc., (3) none of the presented theorems are tested in the experiments, e.g., claims like “greater score differentiation” and “complexity” are neither quantitatively nor qualitatively discussed in the experiments, (4) prompt sensitivity and selection is not discussed at all. This is especially damning for a work focused on improving evaluation.
+  
+[significance] In essence, the work proposes (iterative) ensemble scoring using LLMs. The claim of [line 481] “strong empirical evidence for the effectiveness of the proposed LLM advocate architectures in improving the accuracy of LLM output evaluation” is greatly exaggerated and unsupported. There is good reason to believe that most of the reported improvements over a single LLM-as-judge baseline come from the greatly expanded compute budget and the series of hand-crafted prompts. Similar results might thus be obtained by simply providing a single LLM an expanded compute budget and chain-of-thought style reasoning prompts.
+  
+[clarity] While the presented theorems and proofs in the appendix are an admirable attempt at introducing rigor to LLM-ensemble evaluation. Yet, they also display a limited understanding of the many practical considerations in using LLMs and the large existing literature documenting poorly understood LLM behaviors. Sweeping, unmotivated assumptions like those on line 321, or the assumption that LLMs assigned different “persona prompts” logically will obtain more diverse and stronger arguments limit the usefulness of the presented theorems.
+
+### Questions
+1.  [results] Did the authors analyze the different types of arguments and justifications between the different ensembles in scoring answers?
+2.  [results] Were there any question-answer pairs for which the ensemble methods performed particularly better than the single-judge baseline?
+3. [experiments] How many tokens were needed on average for the different ensembles and models studied?
+4. Section 3.5 is entirely in the appendix, yet referred to in the conclusion [line 499] as discussed. At the minimum, discuss the main results of a section in the main text when referring to it in the conclusion.
+5. [line 504-505] “we have conducted … our framework”: where?
+6. [C.2-C.3] How were any of these chosen? They seem completely arbitrary and unmotivated.
+
+### Soundness
+1
+
+### Presentation
+3
+
+### Contribution
+1
+
+---
+
+## Human Reviewer 3
+
+### Rating
+1
+
+### Rating Number
+1
+
+### Confidence
+5
+
+### Summary
+I suspect that this paper may have been generated by a generative AI (such as ChatGPT). The evidence supporting this suspicion includes:
+
+1. The title of the PDF differs from the title listed on OpenReview.
+2. A significant portion of the literature cited appears to be fabricated. While I have not verified every citation, most of the references listed from 2023 onwards seem likely to be fake."
+
+
+For examples:
+
+[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Eric Wong, Zihang Zhang, Andy Zou, Lianmin Zheng, Siyuan Yu, Yi Tian, Yinghai Zhu, et al. Chatbot arena: Benchmarking open large language models in the wild. arXiv preprint arXiv:2306.01670, 2024.
+
+[25] T. Lanctot, A. Charnock, and J. Badger. Evaluating multi-agent systems in language models. In NeurIPS 2023 Workshop on Multi-Agent Systems, 2023.
+
+[26] Y. Li, D. Chen, and T. Brown. Agents as evaluators: The role of multi-agent systems in llm assessment. In Proceedings of the 2024 Conference on Neural Information Processing Systems (NeurIPS), 2024.
+
+[34] S. M. Panickssery, E. Lee, and K. Lee. Llm-based evaluators for language models: Opportunities and challenges. In Proceedings of the 2024 International Conference on Learning Representations (ICLR), 2024.
+
+### Strengths
+see summary
+
+### Weaknesses
+see summary
+
+### Questions
+see summary
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+---
+
+## Human Reviewer 4
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+The authors propose a framework that interprets Large Language Models (LLMs) as advocates within an ensemble of interacting agents, allowing them to defend their answers and reach conclusions through a judge and jury system.
+
+### Strengths
+The writing is relatively fluent.
+
+### Weaknesses
+- The framework proposed by the authors can be seen as an implementation of a multi-agent approach in the field of LLM-as-judges, with limited novelty and contribution to the community.
+
+- There is a lack of detailed description and justification for the proposed framework, with specific issues highlighted in the Questions section below.
+
+- The authors mentioned probabilistic modeling as one of the key contributions in the abstract (Line 018), but only dedicated a single sentence to this aspect in the main text (Line 395).
+
+- The authors conducted only one experiment, comparing the accuracy of their designed framework with a simple baseline, which is insufficient to support their claims. I suggest that the authors add the following experiments and comparison methods:
+  - **Comparison methods:**
+     - LLMs specifically trained for evaluation, such as PandaLM or Prometheus model.
+     - Multiple LLM evaluators using a majority voting strategy.
+  - **Experiments:**
+     - A comparison of the API and time costs between the proposed MORE and SAMRE frameworks and the aforementioned comparison methods.
+     - A performance comparison of the MORE and SAMRE frameworks under different parameter settings (e.g., the number of advocates).
+     - A bias analysis comparing the MORE and SAMRE frameworks with the aforementioned comparison methods to demonstrate the claim of being "unbiased" (Line 116) and mitigating the influence of strategic behavior and individual biases (Line 119).
+
+### Questions
+- In the proposed MORE framework, why employ three advocates for each answer? Are these advocates different in any way? Additionally, why does judge J provide scores s_1 and s_2 for both answers at the same time (Line 245)? Does this introduce additional bias? I assume the distributions of s_1 and s_2 obtained this way differ from the distributions obtained if s_1 and s_2 were assessed separately.
+
+- What prompts are used for jurors with different backgrounds? I also question whether merely assigning an identity through the prompt (e.g., "A retired professor of ethics") allows the LLM’s evaluation to truly represent the standards of that demographic. This method’s effectiveness requires further validation.
+
+- Could the authors provide an example for the stopping mechanism (Lines 262-263)?
+
+- Why does Algorithm 2 discuss the case of three jurors, when the authors claim five diverse jurors (Line 253)? The authors need to provide the correct version of Algorithm 2.
+
+- Why does the performance of the SAMRE architecture without juries in Table 1 surpass that of SAMRE?
+
+#### Minor Problems
+- The authors should cite reference papers for the theories mentioned in Lines 036-039.
+
+- The authors should clarify the version of the LLMs reported in Table 1. For example, the version of Qwen.
+
+### Soundness
+1
+
+### Presentation
+2
+
+### Contribution
+1

human_reviews/1OP4crhgkD.md

@@ -0,0 +1,206 @@
+# Semantically Aligned Task Decomposition in Multi-Agent Reinforcement Learning
+
+- Decision: Reject
+- Scores: 6, 3, 5, 6
+
+## Abstract
+The difficulty of appropriately assigning credit is particularly heightened in cooperative MARL with sparse reward, due to the concurrent time and structural scales involved.
+Automatic subgoal generation (ASG) has recently emerged as a viable MARL approach inspired by utilizing sub-goals in intrinsically motivated reinforcement learning. 
+However, end-to-end learning of complex task planning from sparse rewards without prior knowledge, undoubtedly requires massive training samples. 
+Moreover, the diversity-promoting nature of existing ASG methods can lead to the ``over-representation'' of sub-goals, generating numerous spurious sub-goals of limited relevance to the actual task reward and thus decreasing the sample efficiency of the algorithm. 
+To address this problem and inspired by the disentangled representation learning, we propose a novel ``disentangled'' decision-making method, \textbf{S}emantically \textbf{A}ligned task decomposition in \textbf{MA}RL (\textbf{SAMA}), that prompts pretrained language models with chain-of-thought that can suggest potential goals, provide suitable goal decomposition and subgoal allocation as well as self-reflection-based replanning.
+Additionally, SAMA incorporates language-grounded RL to train each agent's subgoal-conditioned policy. 
+SAMA demonstrates considerable advantages in sample efficiency compared to state-of-the-art ASG methods, as evidenced by its performance on two challenging sparse-reward tasks, \texttt{Overcooked} and \texttt{MiniRTS}.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This paper proposes the Semantically Aligned task decomposition (SAMA) framework, which aims to solve the sparse reward problem in multi-agent reinforcement learning. SAMA prompts pre-trained language models with chain-of-thought that can suggest potential goals, provide suitable goal decomposition and subgoal allocation as well as self-reflection-based replanning. Each agent's subgoal-conditioned policy is trained by the language-grounded RL method. Compared with the traditional automatic subgoal generation method, SAMA can have higher sample efficiency. This paper verifies the performance of SAMA on Overcooked and MiniRTS.
+
+### Strengths
+1. The paper has a clear structure, introduces the proposed method step by step, and the figures are clear and easy to understand.
+2. Experimental details are given in the appendix, which makes it easy to reproduce the experimental results. Relevant prompts are also given in the appendix, making the contribution and experimental results more convincing.
+3. The testbed chosen in this paper is very representative and challenging. The tasks in Overcooked and MiniRTS can be decomposed by common sense, and their status is relatively easy to translate into natural language. This allows the paper to better focus on how to decompose tasks and allocate subtasks.
+4. It can be seen from the experimental results that SAMA can indeed reach or exceed the performance of existing baselines, and the sample efficiency is indeed significantly higher than other baselines.
+
+### Weaknesses
+1. Currently, SAMA may not be applicable to an environment where human rationality is immaterial or inexpressible in language or when state information is not inherently encoded as a natural language sequence. For example, when the state space is continuous, it is difficult for SAMA to complete the state and action translation stage.
+2. PLM still has some flaws, which sometimes hinder the normal progress of the entire SAMA process. In addition, using PLM will bring additional time costs and economic costs.
+3. In some scenarios (such as Coordination Ring), although SAMA learns very quickly in the early stage, the final convergence results are still not as good as some baselines.
+4. Although in the task manual generation stage, SAMA automatically extracts critical information from the latex file or code through PLM, this is undoubtedly a relatively cumbersome process, so the cost of this stage cannot be ignored.
+
+### Questions
+1. Is the introduction of the self-reflection mechanism unfair to other baselines? Because other methods do not have this ability similar to "regret."
+2. As can be seen from Figures 4 and 5, although SAMA has a very high sample efficiency in the early stages of training, it often converges to a local optimal solution. What is the reason for this result? Is it because the interval $k$ is too large? Do different $k$ values have different effects on the algorithm's performance?
+3. How to ensure the accuracy of PLM in the Task Manual Generation stage or the generated code snippet for assessing sub-goal completion?
+4. Is the wall time for the agent to complete a round in SAMA much different from other baselines? What is the number of tokens that need to be input to PLM in one episode?
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 2
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The authors propose Semantically Aligned task decomposition in MARL (SAMA), a method that aims to generate subgoals for MARL tasks with sparse reward signals. By taking advantage of pretrained language models, the proposed method shows to be more sample efficient during MARL training.
+
+### Strengths
+Generally, the paper is well written and well organized.
+
+This paper:
+* proposes a complex method to integrate language models and MARL from a task decomposition perspective
+* shows that using prior knowledge from language models improves sample efficiency in MARL
+* provides detailed analysis regarding the language learned for the tasks
+
+### Weaknesses
+* Overall, there is a big limitation from the proposed method: the approach introduced requires the prior existence of the required resources to create an accurate task manual and state action translations. This is indeed pointed by the authors in section 3.1: "Nevertheless, linguistic task manuals still need to be made available for general multi-agent tasks. Furthermore, except for a handful of text-based tasks, typical multi-agent environment interfaces cannot provide text-based representations of environmental or agent state information and action information."; this makes this method very limiting
+* It is stated that the language task manual is generated from the latex code of the paper of the multi-agent task. Once again, this sounds very limiting.
+* While the proposed method shows to reduce the required samples in the tasks (Fig. 5), it requires very complex prior preprocessing to create the required text-based rules and manuals. This makes me think that it possible that the method becomes even more costly, in general, after all of this processing, despite the sample efficiency in the task.
+* From my understanding, the goal decomposition is made prior to the task, meaning that a lot of prior knowledge is required (as mentioned before). Other methods such as MASER [1] or [3] decompose the tasks on a more flexible manner, which saves a lot of potential preprocessing.
+* Throughout the paper, the authors claim several times that their method does not need to learn how to split subgoals or to generate them on the go as other methods do, reducing the sample complexity. However, the preprocessing carried needed to achieve this seems very complex and requires a lot of prior knowledge and carefuly engineered features. I wonder again whether this is really more advantageous than following the standard approaches.
+* I have concerns regarding the claims that this method addresses the credit assignment problem. Also since the authors test is environments with only two agents, this can be difficult to analyse (overcooked with 2 agents and MINIRTS with 2 agents; in related works environments with many more agents such as SMAC [2] are used).
+* In the conclusion it is stated as a limitation: "where human rationality is immaterial or inexpressible in language or when state information is not inherently encoded as a natural language sequence."; Yet i believe this is a very interesting remark and would be interesting to see how a method such SAMA can be used to tackle these problems, instead of having a preset of convenient "manuals".
+* The authors mention throughout the paper (introduction and Fig. 2, for example) the potential for generalizability of the proposed method. However, this is not shown or further discussed, and due to the required preprocessing I fail to understand how generalization to different environments/tasks can be easily done.
+
+Overall, I think that the proposed way of integrating language models with MARL from a task decomposition perspective is interesting and can contribute to the explainability of MARL systems. However, I feel that the proposed method in this paper has several limitations and requires a very complex preprocessing. If the manuals cannot be properly generated, then it is not possible to tackle the problems. I also wonder whether the shown sample efficiency is really worth it since it needs all this preprocessing.
+
+[1] https://arxiv.org/abs/2206.10607
+
+[2] https://arxiv.org/abs/1902.04043
+
+[3] https://ieeexplore.ieee.org/document/9119863
+
+
+Minor:
+- in section 2:  "endeavoring to parse it into N distinct sub-goals g1k, · · · , gN"; missing brackets
+- in section 3.1: "illustrate the process in the yellow line of Figure 2" the word figure shouldnt be in yellow; same here (purple line in Figure 2) and in the others that follow in the paper
+
+### Questions
+1. To extend the proposed method to other cases, would it be possible to create a task manual and state action translation for environments that do not follow the conventions presented in this paper? For instance, in more complex environments such as SMAC [2].
+2. In section 3.1: "For each paragraph $S^i_{para}$, we filter paragraphs for relevance and retain only those deemed relevant by at least one prompt from $Q_{rel}$."; how is the filtering of the relevant paragraphs done? Are they manually filtered?
+3. In overcooked, despite the method being more sample efficient during training (figure 5) we can see in figure 4-right that the testing performance of the proposed method stays below other sota methods. Is this because MARL might not be good enough for this environment?
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 3
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This paper investigates long-horizon, sparse reward tasks in cooperative multi-agent RL problems. The proposed method is to use the pre-trained large language model and pre-trained language-ground RL agent. The method prompts the pre-trained language model to generate potential goals, decompose the goal into sub-goals, assign sub-goals to each agent in the multi-agent setting, and replan when the pre-trained RL agents fail to achieve the goal. 
+
+The experiments are conducted on two challenging benchmarks for MARL, Overcooked and MiniRTS. The proposed method outperforms the SOTA baselines.
+
+### Strengths
+The problem of long-horizon, sparse reward tasks in multi-agent RL is super challenging and significant.
+
+It is well-motivated to apply the large-language model, to use prior knowledge and common sense for goal generation and task planning.
+
+Empirically, the proposed method outperforms the SOTA.
+
+### Weaknesses
+In general, the presentation can be improved. Since there are too many components in the pipeline, it will be better to emphasize and explain the most novel and important part in detail, rather than briefly mention each component within space limit.
+
+The proposed method is not fully analyzed. For example, as for the reward design part, how is it accurate to determine task completion? About the pre-trained language-grounded RL agent, how does it perform the training and validation set of states and sub-goals? In the self-reflection phase, how does the task planning evolve?
+
+### Questions
+Could you please clarify in Figure 5, which component of the proposed method is updated as environment steps increase? If the language-grounded RL agent is trained here, why is it called pre-trained?
+
+### Soundness
+2 fair
+
+### Presentation
+2 fair
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 4
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+Based on automatic subgoal generation, the authors design a sophisticated framework to perform goal generation, sub-goal assignment, and language-grounded goal-based MARL, along with effective techniques like in-context learning and self-reflection in the prompting engineering domain. The authors evaluate performance of derived policies on two benchmarks, Overcooked and MiniRTS.
+
+### Strengths
+1. The authors compose several techniques of prompting engineering to realize subgoal generation and train goal-conditioned agents, which seems significantly improve sample efficiency. The authors well introduce multiple advancements from current LLM-agent research and provide thorough discussions about related methods.
+2. The framework the authors propose is comprehensive, though maybe too complicated, it provide a guidance of design an LLM-assisted agent.
+3. The framework, SAMA, realizes semantically useful task decomposition by generating and assign explainable subgoals, which is advantageous compared to previous goal-based methods.
+
+### Weaknesses
+1. The major concern is that the sophisticated design of SAMA may hinder its general use on other benchmarks. It seems that the authors create exhaustive prompts for running SAMA on these two benchmarks (Page 23-36). 
+2. The accompanied concern is that the deployment of SAMA may induce high costs of calling LLMs (like OpenAI API) and training language-grounded agents. 
+3. In the experiments part, the authors do not make some illustrative examples of goal generation but only provide performance curves. 
+
+Some minor mistakes:
+1. In the caption of Figure 2, "subgaol" -> "subgoal". The authors should also consider unifying the term to "subgoal" or "sub-goal" as they both appear in this paper.
+2. Figure 4 (right) is too small to be read.
+
+### Questions
+1. Can you make a list of how many tasks are accomplished with the help of PLMs? Among them, how many can be done in offline setting and how many cannot? I think it will be better to evaluate the contributions. 
+2. To hack SAMA for another environment, can you summarize how many prompts/components the practitioner should modify?
+3. How much percentage of goals and subgoals is generated in the offline manner? How many queries are needed during online training? 
+4. Can you provide the financial costs and time costs of training SAMA agents?
+5. As the evaluation of Overcooked is based on self-play performance, what is the difference among those ad hoc teamwork methods (SP, FCP, and COLE)? In my opinion, the SP may be reduced to a general MAPPO algorithm. Why do ASG methods perform much worse than SP?
+6. Why do you select the MiniRTS benchmark with splitting units rather than directly evaluating on other MARL benchmarks? It seems a little wired. Meanwhile, why cannot SAMA surpass RED?
+
+### Soundness
+3 good
+
+### Presentation
+2 fair
+
+### Contribution
+3 good

human_reviews/3QkzYBSWqL.md

@@ -0,0 +1,189 @@
+# Universal Backdoor Attacks
+
+- Decision: Accept
+- Scores: 6, 5, 5, 6
+
+## Abstract
+Web-scraped datasets are vulnerable to data poisoning, which can be used for backdooring deep image classifiers during training. Since training on large datasets is expensive, a model is trained once and reused many times. Unlike adversarial examples, backdoor attacks often target specific classes rather than any class learned by the model. One might expect that targeting many classes through a naïve composition of attacks vastly increases the number of poison samples. We show this is not necessarily true and more efficient, 
+ _universal_ data poisoning attacks exist that allow controlling misclassifications from any source class into any target class with a slight increase in poison samples. Our idea is to generate triggers with salient characteristics that the model can learn. The triggers we craft exploit a phenomenon we call _inter-class poison transferability_, where learning a trigger from one class makes the model more vulnerable to learning triggers for other classes. We demonstrate the effectiveness and robustness of our universal backdoor attacks by controlling models with up to 6,000 classes while poisoning only 0.15% of the training dataset.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+Whereas in traditional backdoor literature attacks focus on a specific target class, the proposed work introduces a method to embed backdoors from any source class to any target class. The method proceeds in three steps: 1) finding the class-wise centroids of clean-data feature extractions (using CLIP), 2) encoding each centroid into a N-dimensional bit-string, and 3) generating triggers corresponding to each bit-string (and, hence each target class). Classes with similar features are encoded to have similar embeddings. They show that their method performs and scales well with ResNets on four ImageNet-21k subsets.
+
+### Strengths
+- The writing was clear and easy to follow
+- Their bit-string encoding approach is a novel and elegant way to share feature information between classes while generating a class-specific backdoor trigger.
+- The experiments section was well-motivated and well-explained.
+
+### Weaknesses
+- In general, each experiment should be averaged over multiple seeds for statistical significance
+- A major part of the backdoor attack regime is the preservation of clean accuracy, and there is no analysis on how well the proposed method protects a model's clean accuracy. This should certainly be included in future versions of the paper.
+- The proposed triggers in Fig. 2 seem quite obvious to the human eye and may be susceptible to input-space defenses. I would like to see some analysis on the necessary intensity of these triggers and their brittleness to input-space defenses like STRIP.
+- On the defense side, the authors "[halt] any defense that degrades the model’s clean accuracy by more than 2%." I'm open to feedback here, but this has the potential to straw-man some defense mechanisms in scenarios where removing a backdoor is worth the cost of clean accuracy. Including some results without this limitation would be nice.
+- In addition to the above, the attack was not evaluated on data-cleaning defenses like SPECTRE, which I think would be particularly effective against this regime. I would like to see these defenses evaluated as well--and not limited to specific target classes.
+- The experiments are limited to ResNet variants. It would be nice to show generality by including one other architecture in the experiments section.
+  - Since most vision models rely on pretraining, one idea I would find particularly compelling would be to run the attack on a pretrained ViT.
+- In Section 4.4, only a single setting of observed percentage is tried. The analysis here would be stronger if more percentages were tried
+- I'm not sure about the timing here, but the authors claim that they "are the first to study how to target every class in the data poisoning setting." However, while [1,2] address slightly different settings, they seem to be *at least* related and possibly published earlier.
+  - Depending on the nature of this relationship, I would like to see 1) these statements qualified, 2) a more thorough analysis of how the work is positioned in relation to similar work including but not limited to the papers mentioned.
+
+**Citations:**
+
+[1] Du et al., "UOR: Universal Backdoor Attacks on Pre-trained Language Models."
+
+[2] Zhang et al., "Universal backdoor attack on deep neural networks for malware detection."
+
+### Questions
+There are a few questions embedded in the above weaknesses. In addition to those I'm curious about the effect of pretraining on the proposed attack. Could the attack be injected in a fine-tuning regime?
+
+**Note:** I'm happy to raise my score after the weaknesses and questions have been addressed.
+
+### Soundness
+4 excellent
+
+### Presentation
+4 excellent
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 2
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This paper investigates the utilization of a small number of poisoned samples to achieve many-to-many backdoor attacks. The authors leverage inter-class poison transferability and generate triggers with salient characteristics. The proposed method is evaluated on the ImageNet dataset, demonstrating its effectiveness. The authors provide evidence of the transferability of data poisoning across different categories.
+
+### Strengths
+1.The paper demonstrates clear logic.
+2.The topic is intriguing and warrants further exploration.
+
+### Weaknesses
+1.The design motivation of the algorithm is unclear.
+2.The concealment of the patches is poor.
+3.The comparative methods are outdated.
+
+### Questions
+1.	The related work lacks a specific conceptual description of "many-to-many" and an introduction to recent works in this area.
+2.	In Section 3.3, the encoding method used in the latent feature space is rather simplistic, where values greater than the mean are encoded as 1 and others as 0. What is the motivation behind this encoding method, and how does it contribute to improving the transferability of inter-class data poisoning?
+3.	The author employs a patch and blend approach to add triggers, resulting in poor concealment of the backdoor triggers. Visually, the differences between poisoned and clean samples can be distinguished. Has the author considered more covert methods for backdoor implantation, such as injecting triggers in the latent space and decoding them back to the original samples to reduce the dissimilarity between poisoned and clean samples?
+4.	Selection of baselines. The chosen comparative methods are both from 2017. It is recommended to include comparative experiments with the latest backdoor attack methods.
+5.	The experimental results in the paper compare the average attack success rates across all categories. It is suggested to provide individual attack success rates for representative categories or other statistical results such as minimum, maximum, and median values.
+6.	The authors validated the effectiveness of the method under model-side defense measures. It is recommended to include defense methods in data-side.
+
+### Soundness
+2 fair
+
+### Presentation
+3 good
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 3
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This paper introduced a universal backdoor attack, a data poisoning method that targets arbitrary categories. Specifically, the authors crafted triggers by utilizing the principal components of LDA in the latent space of a surrogate classifier. Experiments showed that the generated triggers can attack any category by poisoning a certain percentage of samples in the training data.
+
+### Strengths
+$\bullet$ The authors proposed a method that was designed to poison any class, instead of targeting a single class.
+
+$\bullet$ The proposed attack is effective than the previous method, especially when the poisoning rate is low.
+
+### Weaknesses
+$\bullet$ It is not clear why the proposed method improves the inter-class poison transferability and, in particular, how it ensures that an increase in attack success against one class improves attack success against other classes. Does the proposed method increase the transferability (attack success rate) of any two classes, even if these two classes differ significantly in the latent space?
+
+$\bullet$ The formula in Section 3.2 needs to be formulated more appropriately and clearly. Specifically, do y' and y in the formula refer to any two categories or any two similar categories? If they refer to any two categories, please explain why categories that are very different in the latent space can also improve the success rate of the attack; otherwise, if they refer to any two similar categories, please give a clear definition of similarity.
+
+$\bullet$ The experimental results require further discussion and analysis. For example, in Table 1, the proposed method significantly outperforms the baseline method when the poisoning samples are 5000 (i.e., the attack success rate is 95.5% vs. 2.1%), but the proposed method is suddenly worse than the baseline method when the poisoning samples are 8000 (95.7% vs. 100%). The potential reasons for the sudden improvement in the performance of the baseline method need to be discussed. Similarly, in Table 2, the attack success rate of the baseline method suddenly drops from 99.98% for ImageNet-2K to 0.03% for ImageNet-4K, which also needs to be discussed.
+
+### Questions
+What are the requirements for the surrogate image classifier? The proposed method requires sampling in the latent space of the surrogate image classifier, not the original classifier. Is it possible to use any latent space of any surrogate classifier? For example, if there is a significant difference in the distribution of the hidden spaces between the surrogate classifier and the original classifier, will this result in a significant decrease in the attack success rate of the proposed method?
+
+### Soundness
+3 good
+
+### Presentation
+2 fair
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 4
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper presents a new approach for crafting universal backdoor attacks, i.e. backdoor attacks that target several classes at inference time, as opposed to traditional backdoor attacks that target a single class. In order to mount a universal backdoor attack, the adversary crafts triggers that increase the ASR on several classes simultaneously.  To that end, the authors leverage a pretrained model to extract the feature representation of the training samples, and then craft triggers that correlate with features used by samples from several classes. 
+
+The authors evaluate their attack on several subsets of ImageNet-21k, and against BadNet's baseline presented in Guo et al. By poisoning 0.39% of the training data, the authors are able to mount an effective backdoor attack when no defense is applied. The authors then test the effectiveness of their attack when several defenses are applied, and notice a drop in ASR although the attack remains effective. 
+
+Finally, in order to test how much triggers applied to a single class help triggers applied to other classes, the authors fix the number of triggers in some classes, then vary the number of triggers in other classes, and observe the ASR over the fixed classes increases as more poisoned samples are added to other classes.
+
+### Strengths
+- the paper presents an interesting approach to backdoor attacks where triggers affect several classes simultaneously
+- the authors validate the effectiveness of their attack on a large scale dataset, and against several defenses
+
+### Weaknesses
+- the required number of poisoned samples seems a bit high, even for imagnet. other papers have shown that around 300-500 samples are enough to mount an effective backdoor attack [1, 2]. this is in contrast with the results observed in Table 1, where the baseline attack is not successful even with 2k poisoned samples.
+- the authors only consider a single baseline model against which their attack is compared. this comparison is helpful, however, given the large number of poisoned samples required, it would be nice to see how other baselines would compare at that scale
+- the parameters of the defenses were tuned for a simple baseline (BadNets). the effectiveness of the attack might be very different if the parameters of the defense were tuned to the authors' attack
+
+[1] POISONING AND BACKDOORING CONTRASTIVE LEARNING, Carlini et al., 2022
+[2] WITCHES’ BREW: INDUSTRIAL SCALE DATA POISONING VIA GRADIENT MATCHING, Geiping et al., 2021
+
+### Questions
+- can you please look into a setup with fewer poisoned samples? it should be possible to have a successful backdoor attack with close to 500 samples on ImageNet
+- can you also tune the parameters of the defense against each attack you are considering?
+-  if possible, can you provide a good baseline for attacks to compare against?
+
+### Soundness
+2 fair
+
+### Presentation
+2 fair
+
+### Contribution
+3 good

human_reviews/3flhuT2QGB.md

@@ -0,0 +1,201 @@
+# Towards Synergistic, Generalized, and Efficient Dual-System for Robotic Manipulation
+
+- Decision: Reject
+- Scores: 6, 6, 3, 6
+
+## Abstract
+The increasing demand for versatile robotic systems to operate in diverse and dynamic environments has emphasized the importance of a generalist policy, which leverages a large cross-embodiment data corpus to facilitate broad adaptability and high-level reasoning. 
+However, the generalist would struggle with inefficient inference and cost-expensive training. 
+The specialist policy, instead, is curated for specific domain data and excels at task-level precision with efficiency. Yet, it lacks the generalization capacity for a wide range of applications. 
+Inspired by these observations, we introduce \modelname, a synergistic dual-system that supplements the merits of both generalist and specialist policy. 
+A diffusion transformer-based specialist is devised for multi-step action rollouts, exquisitely conditioned on the high-level task understanding and discretized action output of a vision-language-action (VLA) based generalist. 
+Compared to OpenVLA, \modelname achieves 26.7\% improvement in real-world setting and 12\% gain on CALVIN by introducing a specialist policy with merely 20M trainable parameters. It maintains strong performance with  5\% of demonstration data only, and enables a 3.8$\times$ higher control frequency in real-world deployment. Code would be made publicly available. Our project page is hosted at~\url{https://opendrivelab.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4
+
+### Summary
+This paper investigates a pertinent question in imitation learning: how to combine the generalization capability of models like OpenVLA with the accuracy and task-specific precision of methods such as ACT or Diffusion Policy. To address this, the authors propose a novel framework, RoboDual, which uses the intermediate tokens in OpenVLA to condition a modified Diffusion Policy. Through simulation and real-world experiments, the proposed method demonstrates substantial performance improvements over both generalist and specialist baselines.
+
+### Strengths
+- **Effective Synergy**: The RoboDual framework innovatively combines a generalist and specialist model, leveraging OpenVLA’s generalization and Diffusion Policy's efficiency. This approach addresses a gap in imitation learning by merging generalist adaptability with specialist precision.
+- **Experimental Results**: Both simulation and real-world experiments show significant performance gains over state-of-the-art baselines in generalization and task-specific adaptation, highlighting RoboDual's potential in practical settings.
+- **Well-Written and Accessible**: The paper is clear, well-organized, and easy to understand, making the novel approach and its implications accessible.
+- **Open-Source Commitment**: The authors promise to release the code, which could foster further research and replication.
+
+### Weaknesses
+1. **Limited Real-World Experiments on Generalization** Although the framework shows promise, its real-world experiments, particularly those evaluating generalization capabilities, remain limited. Conducting additional experiments—such as testing with a wider variety of novel objects (beyond just banana to eggplant), introducing more distractors at varied locations, using diverse language instruction templates, varying lighting conditions, or providing more detailed descriptions of the existing experiments—would significantly bolster the case for RoboDual’s practical applicability.
+
+2. **Insufficient Introduction to the CALVIN Dataset** Including illustrative images in Section 4.2 to showcase the training and test settings of the CALVIN dataset would enhance readers' understanding of the experiment RoboDual run in simulation.
+
+3. **Improved Color Differentiation in Bar Charts** The colors representing Octo, OpenVLA, and Ours (single/multi) in the bar figures are difficult to distinguish. Selecting more visually distinct colors would improve clarity and make comparisons easier.
+
+4. **Failure Analysis** It is hard to tell which part is the bottleneck for the current method. A failure analysis will be helpful.
+
+### Questions
+### Questions
+1. From the current experiments, this method does not seem to solve problems beyond what Robot-Moo or RoboPoint achieve. Any thoughts on how this method could outperform approaches based on explicit representations like bounding boxes or points?
+2. It would be easier to understand if the conditioning feature were illustrated in Figure 2.
+3. The authors claim that OpenVLA and Octo serve as generalist models, but they do not generalize effectively in all cases. For instance, the OpenVLA paper mentions challenges with out-of-distribution (OOD) cases, reflective surfaces, unseen action spaces, and actions along the depth axis. Given this, OpenVLA may not be an ideal generalist. Does this imply that RoboDual’s generalization is limited by OpenVLA’s capabilities?
+4. In Figure 3, OpenVLA outperforms RoboDual in the "Knock down object" task. Can the authors explain why this is the case?
+5. Additional real-world experiments on generalization ability would be beneficial. Including a baseline setting where Diffusion Policy/OpenVLA performs reasonably well would also help clarify RoboDual's improvements, given the claim that OpenVLA mainly provides generalization ability in this setup.
+6. Perhaps I missed it, but is the Diffusion Policy baselines in your experiments the modified versions (specialist only), or the originals? This distinction is important, as a significant improvement from switching the transformer backbone to DiT may impact the novelty.
+7. Is the OpenVLA in your experiments the same model used as the generalist in RoboDual (generalist only)?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+---
+
+## Human Reviewer 2
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+3
+
+### Summary
+This paper introduces a new method for solving language-conditioned, robotic manipulation tasks.  The proposed method, RoboDual, combines an vision-language-action (VLA) model for high-level task understanding and long-horizon reasoning with a low-level policy to handle spatial reasoning and precise movements.  The two models are integrated together by passing discretized action predictions and latent representations from the generalist model to the specialist model.  A key benefit of their approach is the ability to run at higher control frequencies (20Hz), which is necessary for many dynamic manipulation skills, since they do not rely on the generalist to make predictions at every time step.  RoboDual is evaluated on a suite of challenging manipulation tasks in simulation and the real-world, where it outperforms all baselines in terms of success rate and shows strong robustness to across task variations.  RoboDual also outperforms baselines even in settings where the amount of data is significantly reduced.
+
+### Strengths
+- The paper proposes a scheme for combining a generalist VLA model with a specialist low-level policy model, via conditioning on latent representations and discretized actions.  This approach relies on the low-level policy to process non-vision-language inputs (like depth, proprioceptive, or tactile info), so the VLA does not need to be fine-tuned extensively.
+- RoboDual can be trained much faster than a fully generalist approach, since the action predictions of an under-trained VLA model are refined by the specialist model that trains quickly.
+- The experiments in Section 4.4 show that RoboDual is very sample efficient, achieving 73.3% success rate at 5 demos on real-world tasks.  This indicates that the coarse action predictions of the generalist are helpful and enable the specialist to refine the actions with limited data.-
+- RoboDual can be run at 15Hz at inference, compared to 4Hz for openVLA.  This difference is significant, since jumpy movements of the robot prevent it from solving tasks that require precision.
+
+### Weaknesses
+- The ablation study needs some work.  Please switch to a different color map so that it is easier to distinguish the bars and read the legend.   The axis range makes it look like the ablations have a substantial impact on model performance, even though the difference in performance is minimal.  There should be error bars, otherwise it is difficult to determine whether a 0.03 increase in average time is significant.  The discussion of these results should also be changed to better reflect the actual results.  For instance, it is not true that "each conditioning source from the generalist model plays an **essential** role in ... enhancing overall performance" [emphasis mine] if removing the conditioning decreases average length by at most 4%.   
+- There are some instances where the wording could be improved.   In Figure 1 caption: "the fast specialist policy *obsesses* ..." (achieves?). Top of page 2, "The *yielding* policy" (The resulting policy).  Bottom of page 2, "We bring in a novel approach" (We introduce? a novel approach).  Beginning of Section 3.3, "Disparate from" (Unlike?).
+- The first contribution says, "Our methodology ... paves the way for the practical application of VLA models".  This is quite a broad claim.  I believe you are hinting at the computational efficiency of the dual-system.  Perhaps modify this to say "practical application of VLA models to higher-frequency control tasks".
+
+### Questions
+- In Table 3, it is interesting that transferring to an unseen background (checkered to solid-white tablecloth) results in 30% or greater drop in performance for all models.  Do you have a hypothesis for why this is the case?  One would expect the generalist models and RoboDual to be more robust to background texture.  
+- What was the reason for choosing joint space control over end-effector control?
+- In Section 3.3, it says that the specialist model is trained for "one hour" but in Section 4.4 it says "one hour of training on a node equipped with eight A100 GPUs".  Is this the same "one hour"?  If so, updating Section 3.3 to "8 gpu-hours" would be more accurate.
+- It seems that fine-tuning the generalist policy to predict discretized actions in a specific robot's action space makes it no longer "generalist".  Have you thought of other ways to condition the low-level policy that might allow one to deploy the system on different robot types?
+
+Typos: Figure 6a legend: (w/o Action L**e**tent).
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 3
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+The paper introduces RoboDual, a dual-system framework combining generalist and specialist policies for robotic manipulation. RoboDual leverages both i) a generalist’s broad generalization capabilities with ii) a specialist’s task-specific precision and efficiency. The generalist is a large-scale pretrained vision-language-action model which provides high-level guidance, while the specialist is a diffusion policy which facilitates rapid, precise control.
+
+### Strengths
+1. Combining a high-level generalist and a low-level specialist model is a compelling paradigm to enable broader generalization while maintaining more fine-grained control.
+
+2. RoboDual achieves higher performance with fewer demonstrations and limited computational requirements.
+
+### Weaknesses
+1. A bi-level policy increases model complexity and therefore inference time, which may affect performance in low-latency tasks.
+
+### Questions
+1. How does the performance vary with different sensory input combinations, and could simpler setups still achieve competitive results while offering advantages in runtime efficiency?
+
+2. How well does RoboDual perform in more dynamic or even user-interactive settings (e.g. moving an object while a trajectory is being executed)?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+---
+
+## Human Reviewer 4
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4
+
+### Summary
+To enable efficient fine-tuning and inference of VLA models, while not compromising generalisability, the article presents RoboDual, a dual-system approach that combines generalist and specialist policies to enhance robotic performance in dynamic environments. The generalist offers adaptability, while the specialist ensures efficient task execution. This synergy results in significant improvements in both real-world and benchmark tasks, offering strong performance with minimal data and higher control efficiency.
+
+### Strengths
+- The figures in the paper are thoughtfully designed and highly informative, significantly aiding readers in understanding the proposed methods and results.
+
+- The dual-system approach aligns well with principles from cognitive science, making its application to embodied manipulation both insightful and innovative. Implementing this concept in robotics is a valuable contribution to the field.
+
+- The extensive experimental results provide strong evidence of the model's advantages in achieving generalizable real-world manipulation. *RoboDual* outperforms both generalist and specialist-only models, demonstrating superior training and data efficiency, which highlights its practical value for broader real-world applications.
+
+### Weaknesses
+- When considering the system from another perspective: viewing the generalist as a vision-language model (VLM) and the specialist model as an action head, RoboDual can be seen as an asynchronous variant of Octo (Ghosh et al., 2024). This diminishes the novelty of the proposed approach, as the heterogeneity between the two systems mainly lies in the data and model scale, which impacts generalizability. It raises the question of whether scaling up the training data for the specialist model might yield comparable performance to the RoboDual system in terms of both computational efficiency and generalizability. Given that DiT is a scalable architecture, and considering the limited dataset used for experiments in Figure 5 (CALVIN), it would be valuable to explore this.
+
+
+    One way to better distinguish these two systems is to draw more deeply on cognitive science concepts, such as viewing one system as responsible for reasoning (akin to System 1) and the other for actuation (akin to System 2). For example, in a task like "write down the answer of 1 + 1 on the blackboard," the reasoning required to determine the answer is challenging for the specialist system alone. Highlighting such distinct roles could provide a more fundamental differentiation between the two systems.
+
+> Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024.
+
+- The experiments on training efficiency could be improved. The use of DistillBERT might not be sufficient for capturing the semantics of actions and target objects from language instructions. I would suggest adding two additional baselines:
+
+    1. Using T5 to encode language for the specialist-only model to ensure sufficient semantic extraction. Since language encoding is performed once per rollout, T5-xxl might be a suitable choice.
+
+    2. GPU hours may not be the best metric for measuring efficiency, as it does not account for the number of parameters. I recommend switching the x-axis metric to FLOPs for a more accurate representation of computational efficiency.
+
+### Questions
+- **Line 26**: Could you clarify why "with" is italicized?
+
+- **Line 268**: The description of the shifted-window conditioning mechanism is somewhat unclear. Why are only $k_g - \tau_s$ generalist actions sampled as conditioning rather than using the entire chunk of $k_g$ actions?
+
+- **Line 197**: There appears to be a duplicated closing parenthesis in ")), \etc". Could you confirm if this is an error?
+
+- In the experiment described in Figure 5, is the generalist model in the dual approach frozen? If it is frozen, are the weights solely from OpenVLA, or has it been further fine-tuned on CALVIN?
+
+- Is the VLA model strictly necessary as the generalist model? If a vision-language model (VLM) were used to extract conditions instead, would this achieve comparable performance to RoboDual?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3

human_reviews/488A64eOf6.md

@@ -0,0 +1,179 @@
+# Language Model Decoding as Direct Metrics Optimization
+
+- Decision: Accept
+- Scores: 6, 6, 5, 8
+
+## Abstract
+Despite the remarkable advances in language modeling, current mainstream decoding methods still struggle to generate texts that align with human texts across different aspects. 
+In particular, sampling-based methods produce less-repetitive texts which are often disjunctive in discourse, while search-based methods maintain topic coherence at the cost of increased repetition. Overall, these methods fall short in achieving holistic alignment across a broad range of aspects. 
+In this work, 
+we frame decoding from a language model as an optimization problem with the goal of strictly matching the expected performance with human texts measured by multiple metrics of desired aspects simultaneously.
+The resulting decoding distribution enjoys an analytical solution that 
+scales the input language model distribution via a sequence-level energy function defined by these metrics.
+And most importantly, we prove that this induced distribution is guaranteed to improve the perplexity on human texts, which suggests a better approximation to the underlying distribution of human texts.
+To facilitate tractable sampling from this globally normalized distribution, we adopt the Sampling-Importance-Resampling technique.
+Experiments on various domains and model scales demonstrate the superiority of our method in metrics alignment with human texts and human evaluation over strong baselines.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+The authors propose a decoding scheme (DAEMON) based on minimizing the divergence of the distribution over generated texts, $q$, with that of the base LM, $p$. They formulate as an optimization problem of this form with a KL objective, provide an approximation, and an efficient sampling scheme for use during decoding.
+
+### Strengths
+- The paper is well written and straightforward to follow. 
+- The motivation is sound, the authors provide an approach for estimating $\mu$, and for sampling 'efficiently' (although the time complexity of this approach is not given).
+- Strong experimental results:
+   - over a range of datasets
+   - against quite a few decoding approaches commonly used in practice
+   - ablation studies
+   - a range of metrics (Repetition, Coherence, Diversity, Information Content)
+- The problem is of great importance in the community.
+
+### Weaknesses
+- Please provide complexity analysis of the sampling approach and compare to competing approaches (Greedy, Top-k, CD, CS, etc). While the experimental results are strong it is important to compare the runtime of this method to determine practical efficacy given the authors claim it is "efficient" in the conclusion. 
+- Analysis on the convergence of $\mu$ in Algorithm 1 and some sensitivity analysis to initialization would be helpful for practitioners.
+- Given that different metrics perform stronger for different # of candidates $M$ and $\tau$, it would be helpful, in terms of robustness, to include some experimental results which show performance on different metrics if one optimizes for a single metric (i.e. coherence).
+
+### Questions
+See weaknesses
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 2
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper introduces a parametric decoding method for generating text from probabilistic models. Essentially, samples from a model are reweighted according to the degree to which they satisfy particular constraints, where the importance of these different constraints is learned via numerical optimization using a small set of samples. These constraints are that the texts generated by the model match human-generated texts according to a variety of chosen metrics. In order to sample from the new (sequence-level) distribution, they make use of a sampling based approximation of the distribution over continuations (of a particular prefix). They show strong results in terms of empirical performance in comparison to standard baseline decoding methods.
+
+### Strengths
+* The method is novel and provides a nice lightweight alternative to fine-tuning methods. It may thus be widely accessible to practitioners without the ability to tune larger language models
+* The empirical component of the paper is comprehensive, including nice ablation studies
+
+### Weaknesses
+* The mathematical motivations given by this paper are quite weakly supported/explained. In general, the language used by the authors with respect to this topic is confusing and informal. For example, they motivate their use of the reverse KL for choosing the parameters of q by stating “KL(q || pθ) restricts the decoding distribution q to deviate minimally from the LM distribution pθ”, but the same argument could be made for the forward variant of the divergence. Similar language is scattered across 2.1
+* The method requires sampling in order to approximate the new parametric distribution (via the WIS algorithm). The runtime of this algorithm, and hence the additional computational complexity incurred by this decoding method, is not discussed. Further, there is an additional sampling + renormalization step that must happen during decoding (algorithm 2). The lack of discussion of these costs is especially pertinent given that the authors point to the tuning required by other decoding methods as one of their downsides “But unlike those previous decoding methods that require heavy manual hyper-parameter tuning for trade-off among different metrics”
+
+### Questions
+* The authors point out that they wish for the difference in the expected value of f under the model and under the data distribution to be 0. It’s unclear to me how this is incorporated into the learning of the different \mu. I see that the target expectation is used in the learning algorithm, but this is not very intuitive. Could some intuition be provided here?
+* “Most existing decoding algorithms lead to deteriorated ε-perplexity, comparing to directly sampling from the input LM distribution.” It seems that [1] shows otherwise 
+* In 2.2, I don’t see why “the second result reveals the perplexity improvement over pθ due to the non-negativity of DKL(pθ,μ || pθ)”
+* The setup for human evaluation is not well explained. What do the numbers in table 3 mean?
+* I do not follow the last statement on page 5 “...so that it can be directly plugged into Eq. (2) to calculate perplexity”
+
+[1] Meister et. al. 2023. On the Efficacy of Sampling Adapters.
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 3
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+The paper proposes a decoding framework for language models that frames it as an optimization problem and performs decoding by optimizing towards desired dimensions. They prove that their method can improve the perplexity on human texts and also experimentally demonstrate that their method is effective on open-ended text generation tasks according to metrics such as repetition, coherence, and diversity.
+
+### Strengths
+1. The proposed framework makes sense and is technically sound. Based on their constructed framework, they propose a reasonable approximation that can work in practice.
+2. Experiments demonstrate good empirical results compared to many other decoding algorithms.
+3. Both automatic and human evaluations are conducted to provide insights into their method.
+
+### Weaknesses
+1. Related works such as [1] are missing. It is worth discussing and comparing with these methods in the paper.
+2. The paper only conducts experiments in two open-ended text generation tasks, where both automatic and human evaluations are hard. On the other hand, evaluations on other text generation tasks such as machine translation and text summarization are more accurate and can better indicate if their method is indeed effective.
+3. Their method can be more computational than baselines.
+
+
+
+
+[1] Fernandes et al., Quality-Aware Decoding for Neural Machine Translation, NAACL 2022.
+
+### Questions
+How efficient is your algorithm?
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 4
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This paper proposes to frame the decoding for natural language generation as a direct metric optimization problem, aiming to generate texts that can align with human texts. To achieves this, the authors select a set of evaluation metric that reflects the quality of generation as constraints and then find the optimal solution to the optimization problem. The authors evaluate their method on a bunch of tasks and report performance in several metrics. According to the results, the proposed method achieves better performance than other decoding strategies such as Greedy, Top-k, and Nucleus decoding.
+
+### Strengths
+- The paper is well-written and easy to follow.
+- The proposed method seems to be intuitive and effective. 
+- The authors provide detailed proof of all the propositions, which is highly appreciated.
+
+### Weaknesses
+- I don't have any big concerns about the paper. However, I have a question regarding the choice of the metric (see questions below).
+
+### Questions
+- Some evaluation metrics might contradict each other. For example, diversity and coherence are hard to achieve at the same time, as far as I know. I am wondering how the results will be if a different set of metrics is used.
+
+### Soundness
+4 excellent
+
+### Presentation
+4 excellent
+
+### Contribution
+3 good

human_reviews/4ihkxIeTFH.md

@@ -0,0 +1,263 @@
+# FAdam: Adam is a natural gradient optimizer using diagonal empirical Fisher information
+
+- Decision: Reject
+- Scores: 3, 3, 3, 3, 3
+
+## Abstract
+This paper establishes a mathematical foundation for the Adam optimizer, elucidating its connection to natural gradient descent through Riemannian and information geometry. 
+We provide an accessible and detailed analysis of the diagonal empirical Fisher information matrix (FIM) in Adam, clarifying all approximations to bridge the gap between the natural gradient and Adam, and advocating for the use of log probability functions as loss, which should be based on discrete distributions, due to the limitations of empirical FIM.
+Our analysis uncovers flaws in the original Adam algorithm, leading to proposed corrections such as enhanced momentum calculations, adjusted bias corrections, adaptive epsilon, and gradient clipping. We refine the weight decay term based on our theoretical framework. Our modified algorithm, \textbf{Fisher Adam (FAdam)}, demonstrates superior performance across diverse domains including LLM, ASR, and VQ-VAE, achieving state-of-the-art results in ASR.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+5
+
+### Summary
+This paper studies the connection between Adam optimizer and natural gradient descent by leveraging techniques from Riemannian geometry. Based on this, the authors propose a modified algorithm named Fisher Adam (FAdam). The convergence analysis of FAdam is provided and the algorithm is tested by large language model (LLM) experiments.
+
+### Strengths
+1. Understanding Adam is an important problem. Using fisher information and natural gradient descent to understand Adam is novel.
+
+2. The presentation of this paper is good in general.
+
+### Weaknesses
+1. The theoretical analysis of FAdam is weak. It directly follows the paper (Defossez et al. 2020) and requires strong assumptions (e.g., $\beta_1$=0, bounded gradient). So it does not analyze the algorithm's momentum and is worse than the state-of-the-art analysis of Adam in the literature. 
+
+2. I do not find any rigorous presentation of Adam's flaws in the paper as claimed in the abstract by the authors. For example, the paper does not have any clear negative results of the vanilla Adam experimentally or theoretically. Therefore, the motivation of designing a new variant of Adam such as FAdam is unclear to me.
+
+3. The description of the experiment is unclear. Lots of details are missing, such as training/test learning curve comparison, learning rate of the optimizer, batch size, and memory costs. Also, the experiment is only run once, and the algorithm's robustness is unclear.
+
+### Questions
+See weaknesses section.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 2
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+The paper reiterates and expands the motivation of Adam as approximate natural gradient descent. It derives multiple modifications to the Adam algorithm based on that interpretation. The resulting method (FAdam) is evaluated experimentally.
+
+### Strengths
+- The argument relating decoupled weight decay to the information-geometric interpretation is interesting. It clarifies that the gradients used to compute $v$ (the diagonal empirical FIM) must be gradients of the log likelihood of a probabilistic model to match the definition of the FIM and therefore must not contain regularizers or auxiliary losses.
+- Averaging the preconditioned gradients (versus preconditioning the averaged gradient) is an interesting variant.
+
+### Weaknesses
+- The paper presents as an original finding that it "establishes a mathematical foundation for the Adam optimizer" in terms of NGD with the empirical Fisher information matrix. This is misleading. This motivation has been given in the original Adam paper and has since been discussed and critiqued in various papers, including but not limited to Kunstner et al. (2019). This should be made transparent in the discussion of related work.
+- The paper states that "for using natural gradient optimizers [...] the loss function must be in the form of the log-likelihood". This is not a factual statement and should be adjusted. Preconditioning with the Fisher information matrix adapts to the geometry induced by a certain probabilistic model. The negative log likelihood under said model may be a "natural" objective function to optimize, but NGD can meaningfully be applied to any other objective. In fact, in Section 3.4.3, the authors advocate for preconditioning an additional loss term with the FIM. 
+- The argument in Section 3.4.1 regarding the use of the square-root on the preconditioner is not stringent. If I am understanding correctly, the argument is that $\Vert \nabla J/\sqrt{f} \Vert^2_2 \approx \Vert F^{-1} \nabla J\Vert_F^2$, i.e., preconditioning with the square-root makes the Euclidean norm of the resulting update equal the  "Fisher norm" of the natural gradient. However, there is no discernible argument why it would be desirable to match these two quantities and, if so, why one would want to achieve this by changing the preconditioner rather than, say, scale the update with a scalar factor? (Minor: The notation should also be improved in Eq. (25) - (27), since $\Vert\cdot\Vert$ is used to refer to both the Euclidean norm and the "Fisher norm".)
+- The paper briefly cites Kunstner et al. (2019), which is an explicit critique of the interpretation of Adam as NGD, but does not really engage with the arguments in that paper.
+- Overall, the paper combines various components, that are somewhat independent of each other:
+    a) introduce gradient clipping,
+    b) apply momentum after preconditioning,
+    c) apply preconditioning to the weight decay gradient.
+It would be highly desirable to perform ablation studies to understand which of these changes actually matter and how they interact.
+- The quality of the empirical evaluation is a bit lacking. No error bars are given. The hyperparameter tuning protocol is somewhat unclear, e.g., FAdam uses a different epsilon value and it is not stated how this value was obtained.
+
+### Questions
+- What is the exact argument for the "invariant natural gradient"?
+- Kunstner et al. (2019) explicitly critique the interpretation of Adam as approximate NGD. What is your response to their arguments? (E.g., degeneracy of the empirical FIM for overparametrized models, sensitivity to model misspecification, no relationship between empirical and true FIM far from an optimum.)
+- How were the $\epsilon$ values in the experiment chose?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 3
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+This work takes a statistical viewpoint of the Adam optiumization algorithm and attempts to both explain it’s performance and add improvlemtents through the lense of the natural graident algorithm. The authors argue that Adam is effectivly preconditioning  with the diagonal of the Fisher Information Matrix,  which leads to  the  algorithm’s superior performance.
+
+### Strengths
+The approach of  analyzing Adam from a statistical  viewpoint is interesting, and while not being new (this interpretation was mentioned in  the origional  Adam paper) it could  deserve a second  look. The authors additionally show some empiracle improvments in a few settings.
+
+### Weaknesses
+While section 2.1 is likely relevant for  doing a detailed  analysis of Adam in the proposed framework, as far  as I can tell that analysis does not actually take place in the paper or appendix. Given this this  section  feels quite out of place to me? I’m unsure of  it’s value for the main message of the paper. Most readers in the optimization or  statistics community are familiar with Fisher Information Matrix based methods and their connections to second order newton style algorithms,  so I’m unsure of the value of introducing them using ideas form differential geometry on manifolds.
+
+Throughout this paper is repeatedly claimed that Adam is preconditioning with the diagonal  of the Fisher Information Matrix, the approximation used in Adam is not the same thing  in general. Adam has been connected  to second order like or natural gradient like algorithms, but it is known in general that the gradient squared is not an approximation to  the diagonal of the Fisher Information Matrix. While it is in expectation, in the finite data regime there  are clear counter examples to this, such as in [1]  which the author cites.
+
+On line  250 the authors claim that Adam excels  in classification tasks such as next  token prediction, but this seems somewhat  contradictory to the previous line where it claimed that CNNs are often better when trained with SGD. I agree that in many vision tasks SGD matches Adam in performance, but what is left out of  the text is that that is true in most classification problems  in vision, which is again a discrete output space. I’m additionally not  sure   of the strength of the claim that Adam is less strong in the generative  setting. I’m aware of works  such as  [2] that claim the oposite (which the author cites) attempting to answer the question “What factors explain that Adam produces better quality solutions than SGDA when training GANs”, and propose modifications to SGD to help it compete with Adam.
+
+Overall several assertions are made that Adam fails on continuous regression targets, but I feel like there  is not sufficient citation or experimentation to back that up. Adam excelling in discrete output spaces (which again is not always true, training  ResNets with SGD is still very common) is not the same thing as Adam failing on continuous tasks, and needs to be justified if it is being claimed. Examples counter to this idea exist in the literature, such as the quadratic function minimized in figure 6 of [3] where Adam handily outperforms gradient descent.
+
+The notation of some of the equations is a but unclear, for example in equation 15 while I understand the division is being coordinate  wise, this should be explicit in the notation,  otherwise a less familiar reader may thing we’re trying to divide a vector by another vector which is ill defined. 
+
+Cosmetically, the citation style is very   non-standard and makes reading difficult, I would suggest the authors use a  more standard method of in text citation.
+
+Minor but Adam Algorithm written in B.5 is in fact AdamW and has clipping added which was  not included in the origional algorithm.
+
+The central weakness of this paper in my opinion is it misunderstands how approximate the approximations in Adam are. The idea of the Adam update being connected to the diagonal of the Fisher information matrix is not new, it was mentioned in Kingma and Ba (2014). The optimization community has tried very hard to understand why Adam works (another weakness of this paper is there is no related work regarding the vast amount of research into understanding Adam) and this approach has not appeared to yield progress. The authors acknowledge the significance of these approximations in appendix B.3, but given the amount of work showing that these approximations are often very poor I don’t think the community can comfortably understand Adam as a natural gradient algorithm.
+
+
+[1]
+Frederik Kunstner, Lukas Balles, Philipp Hennig
+
+Limitations of the Empirical Fisher Approximation for Natural Gradient Descent
+
+https://arxiv.org/abs/1905.12558
+
+[2]
+Samy Jelassi, Arthur Mensch, Gauthier Gidel, Yuanzhi Li\
+
+Adam is no better than normalized SGD: Dissecting how adaptivity improves GAN performance
+
+https://openreview.net/pdf?id=D9SuLzhgK9
+
+[3]
+Frederik Kunstner, Robin Yadav, Alan Milligan, Mark Schmidt, Alberto Bietti
+
+Heavy-tailed class imbalance and why adam outperforms gradient descent on language models
+
+https://arxiv.org/pdf/2402.19449
+
+### Questions
+- Can the authors give a citation regarding using an EMA for fisher info on line 228? I’m not aware that’s been used prior to Adam.
+- How does the proposed framework justify clipping? In B.3 clipping  and epsilon is mentioned  through related work, but this step that has been added to the algorithm does not apppear to be justified by the theoretical framework.
+- What norms are being used in equations (25)-(27)? I’m assuming the first norm is the one induced by the Fisher Information Matrix, but then what is the other one? Euclidian?
+- Has the author tried to quantify how accurate the approximations (F)Adam is using are in a simple setting? This can help figure out if those approximations are in fact reasonable, which needs to be the case in order to claim it’s really natural gradient in disguise.
+
+### Soundness
+1
+
+### Presentation
+2
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 4
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+2
+
+### Summary
+The authors proposed a connection between the Adam optimizer and natural gradient optimization, treating the moving average of squared gradients in Adam as an estimate of the diagonal elements of the Fisher information matrix. They hypothesised that Adam's advantage over other methods might be due to its use of natural gradients; the advantage is particularly noticeable in tasks with discrete distributions, since they allow for a tighter approximation of the Fisher matrix.
+
+The authors also offered a justification for the necessity of normalization by the square root of squared gradients to ensure basis invariance when averaging gradients in Adam. Additionally, they analyzed how momentum, weight decay, and clipping should function in the context of natural gradients and proposed new variants of Adam and Adafactor — FAdam and FAdafactor. The proposed FAdam method demonstrates superior performance for models like LLMs and VQ-VAEs, as well as in ASR tasks.
+
+### Strengths
+1. The authors have identified an important and interesting connection between the success of the Adam optimization method and optimization by the natural gradient method.
+2. The authors proposed an explanation for why Adam's advantages primarily emerge in problems with discrete distributions.
+3. The authors established principles for using momentum, weight decay, and clipping in optimization with invariant gradients. Based on this analysis, they proposed a new method — FAdam, which demonstrates improved performance compared to traditional Adam.
+
+### Weaknesses
+1. The authors did not provide an analysis to assess the accuracy of the approximations and simplifications used in this method.
+  
+a) Why is the transition from sampling from $p(x|\theta)$ to sampling from $p_{data}$ valid? In the case of an undertrained model, the distribution $p(x|\theta)$ can differ significantly from the marginalized $p_{data}$ distribution.     
+  
+b) How accurate is the transition from Eq. (20) to Eq. (21) and why it is not critical to the method's effectiveness?
+  
+c) How accurate is the FIM approximation throughout the hundreds of optimization steps for EMA? 
+
+If the authors could provide ablation experiments comparing Adam, FAdam, a true natural gradient method, and other methods incorporating intermediate transitions on simple tasks (e.g., CIFAR-10), it would significantly increase confidence in the results.
+
+2. In the theoretical justification for preferring discrete distributions: due to uniform sampling from $p_{data}$, discrete distributions can also provide a poor approximation of the FIM. This is because the concentration of the distribution may lie in false logits, which is common in yet-not-fully-trained networks. The score might not be large enough to yield a good approximation.
+
+3. While Amari et al. (2019) prove that unit-wise block diagonal FIM has off-diagonal blocks smaller by $\frac{1}{\sqrt{n}}$, the authors' interpretation appears to extend beyond the original result. Their derived claim about individual diagonal weights dominating off-diagonal weights by $\frac{1}{\sqrt{n}}$ (lines 186-188) may need additional justification, as it's not directly supported by Amari's work.
+
+4. The absence of confidence intervals in the experimental results prevents from being fully certain of FAdam superiority, due to marginal score improvements. Additionally, providing further experiments on a broader range of domains would strengthen the evidence of the proposed method's improved performance.
+
+5. Minor typos: in Eq. (25), it should be the square of the norm; in Eq. (19), "approx." should replace the equality sign.
+
+To summarize, I believe this paper relies too heavily on unjustified approximations and is not yet ready for the conference. However, if the authors provide additional experimental and theoretical validation, I might increase my score.
+
+### Questions
+Please see the weaknesses for questions and improvements.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 5
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+3
+
+### Summary
+In this paper, the authors provide an explanation of the second-order moment $v_t$ of Adam from the perspective of the diagonal Fisher information, and propose a new optimizer, FAdam, by utilizing this perspective.
+
+### Strengths
+This paper provides a comprehensive discussion of the previous works, and the empirical results seem supportive.
+
+### Weaknesses
+At least from my perspective, I can not grasp the main contribution of this paper. I feel it is more like a technical report or a review instead of a paper. I will list my concerns as follows:
+
+1. The main conclusion of this paper is somewhat unclear, and the writing is difficult to follow. From my understanding, the authors aim to claim that Adam is a variant of natural gradient descent and introduce a new optimizer, Fadam, as I outlined in the summary. However, they spend over half of the paper discussing basic statistical properties and formulas related to Fisher information, along with extensive reviews of previous works, without presenting their own results or conclusions. In contrast, the descriptions of the algorithms and the theoretical convergence results are glossed over. While I acknowledge that some discussion of prior works is necessary, I believe it should be integrated with the proposed methods and conclusions of this paper. In summary, the lengthy review of existing literature and preliminary knowledge renders the current manuscript confusing and unappealing.
+
+
+2. The technical contribution of this paper is relatively insufficient. As a work proposing a new optimizer, the authors fail to provide a rigorous theoretical guarantee of convergence. The current version's convergence analysis disregards the effects of momentum, and even this incomplete result is derived directly from another paper. Furthermore, the statement that _“Since FAdam’s momentum is analogous to Polyak momentum, FAdam’s momentum also tightens the convergence bound. Therefore, the convergence bound for the natural gradient without momentum is looser than the convergence bound for FAdam,”_ is presented without adequate justification and is not convincing. It is unreasonable to assert that the convergence bound of one optimizer is looser than that of another without rigorous derivation.
+
+### Questions
+I suggest the authors rethink their major contribution of this paper. As a paper to propose a new optimizer, it might be better to first introduce the new algorithm and present the pros of this algorithm (they could be empirical results of theoretical guarantees). Although I understand some basic illustrations about the preliminary knowledge or motivation of some terms are necessary, at least for this paper, I believe the discussion in the current manuscript should be refined. For example,  I do not get any interesting insights from the discussion about the connection between log-likelihood of Gaussian distribution and $\ell_2$ loss, as it is a basic knowledge of statistics, and seems not deeply correlated with the Fadam.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+1

human_reviews/6IyKniOabO.md

@@ -0,0 +1,141 @@
+# Identifying single molecule force spectroscopy data using deep learning with physics augmentation
+
+- Decision: Reject
+- Scores: 8, 5, 1
+
+## Abstract
+Deciphering the pathways of protein folding and unfolding under tension is essential for deepening our understanding of fundamental biological mechanisms. Such insights offer the potential to develop treatments for a range of incurable and fatal debilitating conditions, including muscular disorders like Duchenne Muscular Dystrophy and neurodegenerative diseases such as Parkinson’s disease. Single molecule force spectroscopy (SMFS) is a powerful technique for investigating forces when domains in proteins fold and unfold. Currently, manual visual inspection remains the primary method for classifying force curves resulting from single proteins; a time-consuming task demanding significant expertise. In this work, we develop a classification strategy to detect measurements arising from single molecules by augmenting deep learning models with the physics of the protein being investigated. We develop a novel physics-based Monte Carlo engine to generate simulated datasets comprising of force curves that originate from a single molecule, multiple molecules, or failed experiments. We show that pre-training deep learning models with the simulated dataset enables high throughput classification of SMFS experimental data with average accuracies of $75.3 \pm 5.3$\% and ROC-AUC of $0.87 \pm 0.05$. Our physics augmentation strategy does not need expensive expert adjudication of the experimental data where models trained using our strategy show up to 25.9\% higher ROC-AUC over the models trained solely on the limited SMFS experimental data. Furthermore, we show that incorporating a small subset of experimental data ($\sim 100$ examples) through transfer learning improves accuracy by 6.8\% and ROC-AUC by 0.06. We have validated our results on three new SMFS experimental datasets. To facilitate further research in this area, we make our datasets available and provide a Python-based toolbox (\url{https://anonymous.4open.science/r/AFM_ML-2B8C}).
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+3
+
+### Summary
+The paper trains a classifier on time vs. force curves from protein unfolding force measurements. For such measurements, the model classifies whether the measurment came from the type of interaction that is desired to be observed, or from other artifacts that should not be included in the assessment of the protein unfolding forces. This has been done before. The application is useful because practitioners can automatically determine which measurments to include in downstream analyses instead of deciding manually. The papers main novelty is a dataset of simulated protein unfolding measurments which are used to train a better classifier next to training on experimental data. The paper is evaluated on three protein's unfolding measurmeents.
+
+### Strengths
+1. The paper introduces an interesting task to the ML conference community (although it is not the first one to address the specific task with ML solutions).
+2. Very good explanations of the new application area.
+3. Good experimental protocols that  ensure statistical significance of the results.
+4. The paper demonstrates that for these measurments, training on synthetic data suffices for obtaining a generalizable classifier that can predict whether force measurements
+
+### Weaknesses
+1. Novelty of the application: The paper applies the most basic ML solution to a task. It can still be valuable to introduce a new task to the ML community. The task has been addressed with similar ML tools in previous works. Thus, there is no novelty in application. There is some technical novelty in the creation of a synthetic dataset.
+2. I am not completely sure whether this is a weakness since I am not familiar with the task: I would imagine that for most proteins of interest we do not have the same sequence and structure repeating multiple times and then observe the same unfolding pattern of essentially the same protein. Why is there no evaluation for proteins that are not repeating or is this the case for one of your 3 experiments? If this is indeed of interest, and we cannot simply combine a single protein of interest into a chain of repeating proteins, then it seems to me that the evaluations miss the important evaluation and only an easier task is evaluated. The task is easier because classifying whether the same unfolding event and the same pattern occurs in the curve multiple times is easy compared to classifying a single unfolding event.
+
+Minor:
+1. The task is very easy. Correct me if I am wrong, but it seems to me that the model could simply classify whether or not there is a repeated pattern in the response measurement.
+
+### Questions
+1. Are the molcules attached to something when we "pull" on them?
+2. As far as I understand, we pull the proteins and observe their unfolding - how are the proteins attached to the head with which we pull?
+3. How do we make sure that he molecules are always attached to the pulling head at the same residue? 
+4. Does the medium in which we measure the pulling force affect the force measurement outcomes? 
+5. Why would we have the same protein repeated multiple times in a chain? Are they attached together in some fashion?
+
+### Soundness
+4
+
+### Presentation
+3
+
+### Contribution
+3
+
+---
+
+## Human Reviewer 2
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+3
+
+### Summary
+The paper presents a method for classifying single-molecule force spectroscopy (SMFS) data using deep learning models enhanced by physics-based simulations. These models classify force curves as originating from no molecule, a single molecule, or multiple molecules, a task traditionally handled through time-consuming expert visual inspection. The authors propose a Monte Carlo simulation framework to create annotated datasets that mimic real-world experimental data, which they use to pre-train deep learning models. The model accuracy improves further with transfer learning on a small subset of experimental data, achieving up to 75.3% accuracy and a ROC-AUC of 0.87 across multiple datasets.
+
+### Strengths
+- Originality
+  - The originality arises from the proposed physics-augmented simulation framework that enables deep learning models to learn realistic representations of SMFS force curves without relying heavily on annotated experimental data.
+- Quality
+  - Extensive comparative analysis across deep learning architectures was conducted.
+  - The proposed method outperforms previous works.
+- Clarity
+  - The method is described clearly.
+- Significance
+  - This study offers a generalizable solution for SMFS data classification, a critical need in the study of protein folding and unfolding mechanisms.
+  - The reduction in dependency on expert visual inspection and annotated data also lowers barriers to adoption, making SMFS analysis more accessible to the broader biological research community.
+
+### Weaknesses
+- Assumptions in simulation
+  - The Monte Carlo (MC) simulation framework assumes identical protein domains across molecules, which could limit the model’s effectiveness for proteins with heterogeneous domains. This assumption may restrict the model’s generalizability, particularly in cases where protein unfolding behavior varies between domains.
+- Computational efficiency and scalability
+  - MC requires considerable computation for each force curve, which might hinder its scalability, particularly in high-throughput SMFS applications that involve thousands of force curves. While this approach is effective in current datasets, its feasibility for larger datasets remains uncertain.
+
+### Questions
+- Impact of Transfer Learning on Model Performance
+  - The paper notes that transfer learning with a subset of experimental data improves model accuracy and ROC-AUC. Could you clarify the minimum amount of experimental data required to achieve significant performance gains?
+- Data Diversity and Experimental Validation
+  - The experimental datasets focus on non-specific pulling for three proteins. Have you considered validating the model on additional proteins or experimental setups (e.g., different solution conditions or functionalized probes)? If not, do you anticipate any limitations when applying this model to other setups?
+- Sensitivity to Simulation Parameters
+  - How sensitive is the model to variations in key simulation parameters, such as those defined by the DHS model? Does the accuracy vary significantly if the estimated parameters are slightly off?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 3
+
+### Rating
+1
+
+### Rating Number
+1
+
+### Confidence
+1
+
+### Summary
+While I recognize the potential significance of the work in understanding protein folding and unfolding using Single Molecule Force Spectroscopy (SMFS) data, I must inform AC and SAC that this specific area is outside my main expertise. Given this, I may not be able to offer detailed or technically relevant feedback that would benefit the review process.
+
+I am happy to provide general comments regarding the manuscript's clarity and structure if necessary, but I wanted to notify you in case you would prefer to reassign this review to someone with more specialized knowledge in SMFS and its related methodologies.
+
+Thank you for your understanding.
+
+### Strengths
+None
+
+### Weaknesses
+None
+
+### Questions
+None
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+4

human_reviews/7M0EzjugaN.md

@@ -0,0 +1,196 @@
+# Online Continual Learning for Interactive Instruction Following Agents
+
+- Decision: Accept
+- Scores: 6, 6, 6, 6
+
+## Abstract
+In learning an embodied agent executing daily tasks via language directives, the literature largely assumes that the agent learns all training data at the beginning.
+    We argue that such a learning scenario is less realistic since a robotic agent is supposed to learn the world continuously as it explores and perceives it.
+    To take a step towards a more realistic embodied agent learning scenario, we propose two continual learning setups for embodied agents; learning new behaviors (Behavior Incremental Learning, Behavior-IL) and new environments (Environment Incremental Learning, Environment-IL) %learning new behaviors (\behaviorILfull, \behaviorIL) and new environments (\environmentILfull, \environmentIL).
+    For the tasks, previous \emph{`data prior'} based continual learning methods maintain logits for the past tasks.
+    However, the stored information is often insufficiently learned information and requires task boundary information, which might not always be available.
+    Here, we propose to update them based on confidence scores without task boundary information \rev{during training} (\ie, \emph{task-free}) in a moving average fashion, named \methodfull (\method).
+    In the proposed \behaviorIL and \environmentIL setups, our simple \method outperforms prior \rev{state of the art} in our empirical validations by noticeable margins.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This paper studies continual learning for instruction-following embodied agents. The authors argue that continual learning is a more realistic setting for embodied AI. To evaluate embodied agents in a continual learning setting, the authors first propose two new continual learning setups for embodied agents:  Behavior Incremental Learning (Behavior-IL) and Environment Incremental Learning (Environment-IL) for new behavior and new environment learning, separately. Then, the authors proposed a new approach for weighting the new logits and old logits. Specifically, a confidence-aware moving average approach which updates logits based on confidence scores of class labels is proposed. The experimental results show that the proposed approach outperforms competitive baselines on the two proposed setups.
+
+### Strengths
+Originality and Significance:    
+The paper is well-motivated. Continual learning for embodied agents is a more realistic yet less explored area in our community. In addition, the two new continual learning setup seems reasonable and could be valuable for evaluating continual learning agents. 
+
+
+Quality:   
+The paper is technically sound. Most of the claims are supported by experimental results. 
+
+
+Clarity:    
+The paper is generally well-organized and implementation details are presented. However, some technical details may need further clarification. Please see the Weakness section for more details.
+
+### Weaknesses
+1. While the empirical results seem promising, the reviewer is not fully convinced by the proposed confident-aware moving average approach. Specifically, it is unclear why the average of most recent N confidence scores could be a good indicator of the ‘quality’ of the newly obtained logits. More concretely, the behavior $t_i$ of the current episode is different from the behavior $t_j$ that is associated with the sampled past experience $x’$. Why would confidence scores of current episodes with behavior $t_i$ provide information on a different behavior $t_j$.   
+
+
+2. In equation (1), subscript $a, c$ are missing in $\gamma_a$ and $\gamma_c$
+
+### Questions
+1. Please elaborate why the confidence scores are good indicators of the ‘quality’ of the new logits. Giving simple examples that demonstrate the effectiveness could be very helpful.   
+
+2. This paper uses imitation learning to train the agents. Is the proposed approach extendable to a reinforcement learning setting?
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 2
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This work looks at the continuous learning problem for embodied agents that follow verbal task instructions. The outlined challenge is very relevant in the community as such agents need to constantly adapt to new tasks and environment challenges. In this work, the authors address Behavior-IL and Environment-IL through a newly proposed method, Confidence-Aware Moving Average, that updates an experience replay of stored logits based on model confidence. The proposed method is evaluated over a set of tasks and behaviors in the ALFRED simulation environment against state-of-the-art baselines.
+
+### Strengths
+* The proposed work addresses a relevant challenge for the deployment of embodied agents in challenging home environments, particularly addressing the problem of adapting previously learned policies to new environments and tasks.
+* The idea of using confidence to update the stored logits is interesting, and potentially quite powerful (if there is any way to know that the confidence is reasonable and/or calibrated, see below).
+
+### Weaknesses
+* I think some of the assumptions this paper makes are not realistic. E.g., the paper indicates that this approach is task-free and that task-identifiers are not available, but then appears to assume that the data is balanced. In a real-world continual learning setup there is no guarantee that the tasks/data the agent encounters will be balanced, and so if data balance is enforced via sub-sampling I think this is implicitly leaking information regarding task identifiers (as they would be needed to balance an unbalanced dataset in practice). An analysis on how class/environment/behavior imbalance affects the model would be beneficial here.
+* The performance improvements seem relatively marginal for many of the evaluation metrics. Were any statistical significance results run over these values?
+* The training methodology is not entirely clear to me, and I think some of it should be pulled into the main paper as it reduces the clarity of the empirical results. Are all the models trained sequentially over a set of tasks/environments and then evaluated over the same set (in the seen case)? Do the seen/unseen evaluations use the same underlying trained model?
+
+### Questions
+1) Is there a concern about over/under-confidence with such an approach? Presumably, such a model is not calibrated since the model is continually out-of-distribution. Given that modern deep neural networks are particularly susceptible to confidence issues, an analysis on this aspect would be useful.
+2) Does performance depend on the order of the tasks/environments? I see reference to random ordered sequences in Sec. B.3, but it's not clear to me how this ties into the training/evaluation methodology.
+3) Is there any insight on how the size of the memory buffer relates to the number of learning tasks? Is 500 an arbitrary number or was this empirically chosen?
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 3
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper studies the online continual learning problem without task boundary information in an imitation learning context. It adapts a distillation based approach that was developed for the setting with task boundaries. Prior methods store past model logits in memory and update these with a weighting scheme to prevent these logits of becoming outdated using task boundary information. Here the weights are determined dynamically via the confidence of the past N predictions of the model, not relying on task boundary information. The method shows improved task performance compared to a variety of continual learning baselines in a simulated household environment. The evaluation is split into a setting where the agents has to continually learn new behaviors and a setting where the agent has to continually adapt to new environments.
+
+### Strengths
+- The way the method and experimental setup is presented, it can be easily followed and understood. 
+
+- The method is tested against a variety of baselines representing different approaches to the continual learning problem.
+
+- The method outperforms all baselines.
+
+- Although the novelty is on the smaller side it becomes apparent how the work relates to prior approaches and what limitation is tackled. It is also shown that tackling exactly that limitation leads to better performance.
+
+### Weaknesses
+- Depending on the availability of compute it would be beneficial to increase the number of seeds (3->5) to reduce the standard errors of the results. At the moment the standard errors are often larger than the gaps in mean performance. 
+
+- Effect of the hyperparameter N (number of past confidences averaged over). I think it would be interesting and important to see how sensitive the method is to that hyperparameter. If it works only well for a specific N, a lot of hyperparameter tuning will be needed to find that N for a given environment. If performance is robust with respect to N it would make the method more easily applicable. 
+
+- The evaluation is split between environment changes and behavioral changes. I understand the point of separating these two factors, but why was the crossing of the factors not studied? What happens if two factors change at the same time?
+
+- As shown in Figure 3, when a new tasks arrives confidence drops. However, there is also evidence that in OOD cases a model can be confidently wrong (Do Deep Generative Models Know What They Don't Know?, Nalisnick et. al.). How detrimental would that be for performance?
+
+Formalities:
+
+- Abstract, last sentence: "outperforms prior arts", was that supposed to be "outperforms prior state of the art" ?  
+- Figure 4: In the second paragraph of 5.3 that references the figure 4, "HEAT" is claimed to be the newly learned task and Pick2&Place the old one, while in the caption of the figure it is the other way around.
+
+### Questions
+- The performance of the "finetune" baseline with respect to the $SR_{last}$ and $GC_{last}$ metrics. If I understand the baseline correctly is that the currently trained policy is finetuned on the data of the next task without any memory. Should the performance on the final task not be . Why is the performance on the . 
+
+- Which episodes are stored in the memory M? Is it simply a queue? If yes then given that its capacity is such a small percentage of the total number of episodes (as said in the text that is the standard) and given the number of tasks for each of the evaluation episodes will most episodes not be about the same same task such that the logits there should be replaced with the current logits of the model?
+
+- In the CAMA w/o D.C. baseline, was the fixed coefficient tuned for optimal performance of the baseline?
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 4
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This paper first introduced two new settings to coninual learning embodied AI: behavior incremental learning and environmental incremental learning. The paper then introduces a method to tackle both, in a task boundariless setting, called Confidence-Aware Moving Average. The method is employed in Alfworld, and the task at each step is to predict action and object class.
+
+### Strengths
+- The continual learning setting is a relevant topic to be explored in the embodied AI field
+- the writing is clear and the method is well motivated and easy to understand
+- the number of baselines compared is very comprehensive
+- the method is simple yet effective
+
+### Weaknesses
+- conceptually, the task boundariless scenario in embodied instruction following AI is very limited, since the user will know when to switch to a new task
+- introduction to a new hyperparameter for each of the prediction ($\alpha$'s)
+- another baseline I could think of is to use the loss as the weight to dynamically change the old logits. What are some intuition on why confidence is better than loss to use to dynamically weight?
+
+### Questions
+- equation (1), the $\gamma$  in front of $z$ is it missing the subscript $a$ and $c$? or is it a fixed value?
+- for equation (2), how are the confidence scores calculated, mathematically
+- I am not too familiar with the benchmark, but when is the object class used when making decisions? if you need to heat a mug, does your agent also need to recognize the object in front of it is a mug too?
+
+### Soundness
+3 good
+
+### Presentation
+2 fair
+
+### Contribution
+2 fair

human_reviews/A8et2yjbly.md

@@ -0,0 +1,124 @@
+# Cross-Modality Masked Pre-training for Visible-Infrared Person Re-identification
+
+- Decision: Reject
+- Scores: 3, 5, 1
+
+## Abstract
+Visible-Infrared person re-identification is a challenging yet important task in the field of intelligent surveillance. Most existing approaches focus on designing powerful deep networks to  learn modality-shared representations, while little attention has been paid to using pre-training methods, although they can improve the performance of cross-modality tasks stably. This paper proposes a cross-modality masked pre-training (CMMP) method for visible-infrared person re-identification. Specifically, we generate color-irrelevant images using random channel exchangeable augmentation to minimize the difference between modalities at first. In the pre-training process, the visible together with the generated image, and the infrared image are masked by sharing the same random mask. Considering the misalignment of visible and infrared images in the datasets, we then reconstruct the masked areas only of the visible and the generated images using a lightweight decoder, which makes the pre-training process more efficient. Extensive experiments on two visible-infrared person re-identification datasets verify the effectiveness of the proposed method. CMMP outperforms the baseline method by +1.87\% and +1.24\% mAP on SYSU-MM01 and RegDB, respectively.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.
+
+### Summary
+This paper presents a mask-based pretraining strategy for visible-infrared person re-identification. Based on the extended images via random channel exchange, the proposed CMMP employs a mask-sharing mechanism to simultaneously mask three types of images. Subsequently, the model is encouraged to reconstruct the masked regions via a lightweight decoder.
+
+### Strengths
+Good writing: I can easily catch the main story of this paper, and the organization of the paper is clear.
+
+### Weaknesses
+1.	Limited novelty: I can hardly catch the significant contribution and inspiring insights for this field. Since both the model architecture and data generation method in the paper are based on existing methods, thus, what is the core value or unique contribution of this paper?
+2.	Lack of citations to recent literature: This paper lacks citations to recent literature, with only one paper from 2021 mentioned in the related work section.
+3.	Outdated experimental comparisons: The experimental results lack comparisons to the latest work. Since this is a submission to ICLR 2024, it is inappropriate to rely solely on experimental comparisons with those prior to 2021. 
+4.	Inaccurate statement about SimMIM: The statement that SimMIM only encodes the unmasked parts is incorrect. SimMIM also encodes the learnable mask features. 
+5.	Lack of clarity in explanations: The paper repeatedly mentions "misalignment inherent in the VI-ReID dataset" without providing an explanation of the negative impact it causes or whether any existing work has addressed this issue. Additionally, the descriptions of pretraining and finetuning processes are not clear. It is essential to provide clear and detailed explanations to enhance the understanding of the proposed method.
+6.	It is not much of a contribution for firstly introducing the mask reconstruction paradigm into VI-ReID. Clarifying the rationality and necessity of introducing it into the task is more important and meaningful.
+7.	The and writing and reference format are sometime irregular, such as ‘Ye et al.Ye et al. (2021)’ in Section 2.2, ‘10 experiments’ in Section 4.1 and ‘288*144 size’ in Section 4.2.
+8.	For the ablation study on the masking strategy, the experimental comparisons with more masking strategies should be provided, rather than only the random mask.
+9.	There is a lack of insight into the experimental results, and the view is superficial analysis on results.
+
+### Questions
+1.	Both MAE and SimMIM, the most relevant models to this paper, adopt the ViT as the backbones. However, the proposed CMMP model uses a CNN architecture. Could the authors provide the reason for this choice? Did you try to apply CMMP on the ViT?
+2.	Overall, the encoding strategy of CMMP seems more similar to SimMIM. Have the authors attempted the MAE-like strategy that only encodes the unmasked parts? Given that CMMP has an 80% mask rate, using the MAE-like encoding could significantly improve training efficiency.
+3.	The paper mentions the statement multiple times -- “Due to the misalignment inherent in the VI-ReID dataset, CMMP focuses its reconstruction efforts solely on the visible image and the generated images after masking”. Could the authors explain what this misalignment refers to and why reconstruction is not performed on the infrared image?
+4.	Could the authors provide a more detailed explanation of the finetuning process? Specifically, what are the differences between pretraining and finetuning in this paper?
+5.	The encoder processing flow should be explained more thoroughly. Figure 2 shows three input images, while Figure 3 only depicts two input images to the encoder. This discrepancy raises concerns about inconsistency. Please clarify this discrepancy and ensure the consistency of the depiction.
+6.	It is recommendable to include more figures in the experimental section. After all, compared to the tables, graphical presentation can provide a more intuitive and direct comparison.
+
+### Soundness
+2 fair
+
+### Presentation
+3 good
+
+### Contribution
+1 poor
+
+---
+
+## Human Reviewer 2
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.
+
+### Summary
+Existing visible-infrared person re-identification(VI-ReID) methods often overlook the disparity between pre-trained data (ImageNet) and VI-ReID data. This paper introduces a novel approach called Cross-Modal Masking Pre-Training (CMMP) for the recognition of visible and infrared humans. Notably, performance enhancements were realized through an initial pre-training stage involving masking and reconstruction, followed by fine-tuning.
+
+### Strengths
+1. This paper introduces a novel approach to the VI-ReID task, focusing on model pre-training for the first time.
+2. This paper has developed a lightweight pre-training method and demonstrated its effectiveness in enhancing VI-ReID.
+
+### Weaknesses
+While the focus of this paper is innovative, the experimental results indicate a weakness in its performance.
+
+### Questions
+1. In this paper, the pre-training data and fine-tuning data are identical. If we conduct pre-training based on ImageNet and then proceed with fine-tuning, does this approach offer any performance advantages compared to the optimal method?
+2.This paper mentions “When attempting to predict all three cross-modality images simultaneously, the model’s performance is notably worse, considerably lower than when predicting other targets.” Does this observation suggest that pre-training doesn't necessarily require infrared image data, and instead, we could use visible and generated images from ImageNet datasets or large ReID datasets for the pre-training process?
+
+### Soundness
+2 fair
+
+### Presentation
+2 fair
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 3
+
+### Rating
+1
+
+### Rating Number
+1
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper includes AUTHOR CONTRIBUTIONS and ACKNOWLEDGMENTS sections in the main text. Therefore, this paper violates the double-blind reviewing rules.
+
+### Strengths
+-
+
+### Weaknesses
+-
+
+### Questions
+-
+
+### Soundness
+1 poor
+
+### Presentation
+1 poor
+
+### Contribution
+1 poor

human_reviews/BrjLHbqiYs.md

@@ -0,0 +1,142 @@
+# Multimodal Learning Without Labeled Multimodal Data: Guarantees and Applications
+
+- Decision: Accept
+- Scores: 6, 6, 8
+
+## Abstract
+\vspace{-2mm}
+    In many machine learning systems that jointly learn from multiple modalities, a core research question is to understand the nature of \textit{multimodal interactions}: how modalities combine to provide new task-relevant information that was not present in either alone. 
+    We study this challenge of interaction quantification in a semi-supervised setting with only labeled unimodal data and naturally co-occurring multimodal data (e.g., unlabeled images and captions, video and corresponding audio) but when labeling them is time-consuming.
+    Using a precise information-theoretic definition of interactions, our key contribution is the derivation of lower and upper bounds to quantify the amount of multimodal interactions in this semi-supervised setting.
+    We propose two lower bounds: one based on the \textit{shared information} between modalities and the other based on \textit{disagreement} between separately trained unimodal classifiers, and derive an upper bound through connections to approximate algorithms for \textit{min-entropy couplings}. We validate these estimated bounds and show how they accurately track true interactions. Finally, we show how these theoretical results can be used to estimate multimodal model performance, guide data collection, and select appropriate multimodal models for various tasks.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper explores the challenge of understanding how different modalities combine to provide new task-relevant information in a semi-supervised multimodal learning setting. The authors derive lower and upper bounds to quantify the amount of multimodal interactions, with the lower bounds based on shared information and disagreement between unimodal classifiers, and the upper bound derived through connections to min-entropy couplings. The authors also propose a practical algorithm to estimate the lower bounds in practice. The theoretical results in this paper could be used to guide data collection or select appropriate multimodal models for a specific task. Overall, this paper provides a valuable contribution to the field of multimodal learning by providing a framework for quantifying interactions between modalities.
+
+### Strengths
+1. The paper provides a novel contribution to the field of multimodal learning by deriving lower and upper bounds to quantify the amount of multimodal interactions in a semi-supervised setting. This provides a framework for understanding how different modalities combine to provide new task-relevant information.
+
+2. The paper is theoretically rigorous, with the authors providing a detailed information-theoretic analysis of the problem. The authors also propose a practical algorithm to estimate the lower bounds in practice. The authors show that synergy bounds can be used to predict the performance of multimodal models on held-out data, and to identify modalities that are most important for synergy. This information can be used to guide data collection efforts and improve the performance of multimodal models.
+
+3. The theoretical results in this paper could be used to guide data collection or select appropriate multimodal models for a specific task. This has important practical implications for the development of multimodal learning systems in a variety of domains.
+
+### Weaknesses
+1. The synergy bounds are only approximate. The authors acknowledge that the synergy bounds are not tight, and that they may underestimate or overestimate the true amount of synergy in a dataset. This is a limitation of any information-theoretic approach to synergy quantification. While the paper provides a theoretical framework for quantifying interactions between modalities, there is limited empirical evaluation of the proposed approach. The authors only provide a proof-of-concept experiment on a small dataset, which may limit the generalizability of the results.
+
+2. The synergy bounds are sensitive to the choice of auxiliary distributions. The authors use a set of auxiliary distributions to compute their lower bound on synergy. The choice of these auxiliary distributions can affect the tightness of the bound. The authors do not provide any guidance on how to choose the auxiliary distributions, which leaves this as a practical challenge for users of the proposed method.
+
+3. The synergy bounds are not applicable to all types of synergy. The authors focus on quantifying synergy that arises from the disagreement between unimodal predictors. However, there are other types of synergy, such as synergy that arises from the complementary strengths of different modalities. The proposed synergy bounds are not applicable to these other types of synergy.
+
+4. The paper does not provide any theoretical guarantees on the performance of multimodal models trained using synergy bounds. The authors show that synergy bounds can be used to predict the performance of multimodal models on held-out data. However, they do not provide any theoretical guarantees on the accuracy of these predictions. This is a limitation of the paper, as it would be useful to know how accurately synergy bounds can predict the performance of multimodal models in practice.
+
+### Questions
+1. In section 3, what is D_M? Did you consider number of modalities over 2?
+
+2. In Table 3, can you give more details of multimodal or unimodal, and what are the evaluation metrics to define the best models?
+
+3. In section 4.2 RQ1, what performance to estimate? I think authors could have provided more explanations or examples to make the paper more accessible to a wider audience.
+
+4. The authors mentioned observing only labeled unimodal data and some unlabeled multimodal data, did authors conduct experiments on these kind of data or any ablation studies on incompleteness of unimodal data?
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 2
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This paper provides two lower bounds and one upper bound to quantify the interactions between modalities for the proposed multimodal semi-supervised setting. One lower bound is based on mutual information between modalities and the other is based on disagreement between separately trained unimodal classifiers. The upper bound is approximated by min-entropy coupling methods. The performances of the bounds to track true interactions on synthetic and real-world datasets are evaluated by experimental results. Examples are shown to use the bounds to guide modality or model selection in multimodal learning.
+
+### Strengths
+1. The performances of the lower bounds are tight, which demonstrates the effectiveness.
+2. The proposed bounds can be used for modality or model selection, which are helpful for multimodal learning.
+
+### Weaknesses
+1. The performance of the proposed upper bound is not very ideal, more explanations can be provided.
+2. No related work is compared in the experimental results. The superiority of this work is not demonstrated.
+
+### Questions
+See the weaknesses.
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 3
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This study develop a system of information-theoretic definitions and theories that quantify multimodal interaction in a semi-supervised setting. The semi-supervised setting being investigated contains only labelled unimodal datasets and unlabelled co-occurring multimodal data. The authors derive 2 lower bounds and 1 upper bound for Synergy information - emerged information when all modalities incorporate. Additionally, the research demonstrates how these theoretical results can be applied to estimate multimodal model performance, guide data collection, and select appropriate multimodal models for various tasks.
+
+### Strengths
+Contribution:
+
+- The definitions are set out in accordance with real scenarios and datasets (e.g. MIMIC, MUSTARD,…) and in accordance with previous literatures, which increase the credit and sensibility of the definitions.
+- The line of theorem following and based on those definitions are set out reasonable and come with rigorous proof in Appendix.
+- By extracting 4 information factors given pairwise marginal distributions, the authors derive the bound of accuracy for optimal multimodal model. This can be used to analyse whether to collect more data and choose a different fusion approach
+
+Prepresentation: 
+
+- The authors provides trackable notation system and use those notations throughout the manuscript, which increase its consistency and coherence.
+- Clear instructions via reference system and details in Appendix, make it more readable and trackable (e.g. we refer reader to …).
+
+### Weaknesses
+- Upon deriving lower bound using uniqueness, the reliance on optimal unimodal classifiers seem not practical in most scenarios.
+- Upon deriving the upper bound, the relaxation using min-entropy coupling problem might not produce a tight upper boundary.
+- While the empirical result support the correct boundary of *Synergy*, there is no investigation on how the bound is closed to *Synergy*, or what is the gap can be determined as close enough.
+
+### Questions
+How to determine the closeness of the derived boundary?
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good

human_reviews/Cn5Z0MUPZT.md

@@ -0,0 +1,206 @@
+# Process Supervision-Guided Policy Optimization for Code Generation
+
+- Decision: Reject
+- Scores: 8, 6, 3, 3
+
+## Abstract
+Reinforcement learning (RL) with unit test feedback has enhanced large language models’ (LLMs) code generation, but relies on sparse rewards provided only after complete code evaluation, limiting learning efficiency and incremental improvements. When generated code fails all unit tests, no learning signal is received, hindering progress on complex tasks. To address this, we propose a Process Reward Model (PRM) that delivers dense, line-level feedback on code correctness during generation, mimicking human code refinement and providing immediate guidance. We explore various strategies for training PRMs and integrating them into the RL framework, finding that using PRMs both as dense rewards and for value function initialization significantly boosts performance. Our approach increases our in-house LLM’s pass rate from 28.2\% to 29.8\% on LiveCodeBench and from 31.8\% to 35.8\% on our internal benchmark. %Additionally, PRMs enhance code generation in long-horizon scenarios. 
+Our experimental results highlight the effectiveness of PRMs in enhancing RL-driven code generation, especially for long-horizon scenarios.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+3
+
+### Summary
+The paper introduces a practical approach for training models with process-oriented feedback for code generation. The approach uses a novel automatic (LLM + unit test -based) data generation process to create a dataset of code with quality labels on every line. The line-by-line labels can be used to trained a Process Reward Model (PRM), which can in turn be used as a dense reward signal for training code models with RL. The approach is shown to outperform the baseline RL-trained code models which are trained on sparse rewards from unit tests. The authors perform careful experiments and ablations to motivate each part of their design.
+
+### Strengths
+Nice work overall! I enjoyed reading this and have become more bullish on PRMs as a general research direction after this.
+- Very readable and easy to understand.
+- Clearly a lot of work was put into it: Devises a novel practical approach for training models with PRMs for code generation, with careful work put into ablations and studying each component.
+- Competitive with the SOTA RLTF approach
+- Clever method to generate process-level supervision data to train PRMs! The idea of using a best-of-K oracle to label whether each code prefix is feasible or contains unrecoverable errors is a non-obvious but effective way to generate data.
+- Section 4.2 and 4.3 were great! Well-reasoned experiments and execution, with great experiment-backed insights on how best to use PRMs in this domain.
+  - Particularly appreciated the detail-to-attention uncovering PRM Hacking and implementing mitigations.
+  - I like that the main results (Table 1) independently shows the effect of introducing [Value Init], [Dense Reward], and [Value Init + Dense Reward]! Very clear.
+
+### Weaknesses
+No clear weaknesses come to mind. There were some choices made which I had different ideas about, but this is not a critique of the work or the claims, so I have put those into the Questions section.
+
+### Questions
+- > To ensure that the PRM training data effectively covers the state space the language model may encounter during the next RL training phase, we sample policy models from various stages of the RL baseline training. Specifically, we select 4 checkpoints evenly spaced throughout the RL baseline model’s training process. For each checkpoint, we sample n responses for each coding prompt in the training dataset Dtrain. For each sampled response, we apply the binary search labeling procedure described in Algorithm 1
+
+  For the oracle, is this also using the policy model checkpoints? I see "Our method leverages the model’s own capabilities to generate completions for partial code prefixes and uses automated testing to assess their correctness" which suggests that the policy model itself is used to generate the best-of-K samples. I can appreciate that not requiring a separate oracle model is nice because it is self-contained, but I think this will result in a suboptimal dataset compared to e.g. using a fully-trained RLTF code model as the oracle.
+
+- Reading Section 4.2.2 on RL Training makes me think: Wouldn't a different scheme of data labelling, which labels each line with the "marginal contribution of the line toward success" be more effective than simply rating [0: infeasible, 1: feasible]? Specifically, my intuition is that each added line should improve the success rate of the oracle given K attempts, so my proposed reward is something like "the reward for step M should be the `(success_rate_of_oracle_at_step_M - success_rate_of_oracle_at_step_Mminus1)`". This captures the idea that each line should increase the likelihood of the program succeeding, and naturally avoids reward hacking by simply adding more lines. Of course, this is easier said than done, I expect the process to be noisy, but this formulation for the dataset seems to be better aligned to the true objective.
+
+- (Not a weakness, just a typo) In Section 3.2, I think the authors meant to do citep instead of just cite: "In mathematical domains, LLMs may generate correct answers with faulty reasoning Lightman et al. (2023), making intermediate verification essential" and "While preliminary attempts have been made to incorporate PRMs into RL trainingWang et al. (2024a)"
+
+### Soundness
+4
+
+### Presentation
+3
+
+### Contribution
+3
+
+---
+
+## Human Reviewer 2
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4
+
+### Summary
+This paper proposes integrating the Process-Reward Model (PRM) that provides line-level rewards into PPO training for single-turn code generation. The authors design a binary search procedure to gather line-level rewards, which is defined as being able to lead to correct code (the authors use best-of-K sampling to approximate it) for PRM training. The authors propose using the PRM as a source of dense reward and/or value function initialization and evaluating the methods' effect on LiveCodeBench and the in-house benchmark using the in-house model.
+
+### Strengths
+The contributions of the paper are the following:
+1. Adapting the PRM to the PPO pipeline for code generation
+2. Details of how the authors design the binary search, collect, and choose PRM training data
+3. Empirical results on LiveCodeBench and in-house benchmark show the effectiveness of PRM and analysis to investigate the source of gain
+
+The paper is well-written and easy to follow. It also highlights challenges when incorporating the PRM into the PPO framework (such as reward hacking in Sec 4.3.3). To mitigate reward hacking, it proposes length normalization and neutral labels for comment lines.
+
+### Weaknesses
+1. My primary concerns are reproducibility, the lack of experimental details, and data transparency, which hinder the community from reproducing the results presented in the work and accessing the needed efforts. This paper uses an in-house model, and part of the results are reported on the in-house benchmark. This is not inherently an issue but means that there needs to be more details in the text/contributions in other components. Therefore, I encourage the authors to include the following:
+-----
+#### Description of the In-house model
+1. What is the parameter size and architecture? 
+2. What is the size of the SFT set, if available (number of tokens/data points) mentioned in L241 and L337? Could the authors show an example of the SFT data? Could the authors provide some basic configuration of the SFT (e.g., lr, number of gradient steps)?
+
+#### PRM data and training
+1. Could the authors give some statistics on the different strategies of PRM training data, if available (number of tokens/data points, avg. lines of code, the distribution of the % or the line number where the line-level label turns -1 from +1)? These statistics could be a valuable add-on to Table 2 and the plain strategy. Could the authors show one example of the reward produced by the binary search?
+
+#### PPO training
+1. what is the KL penalty (the value of $\beta$ in Eq(3)) used in the exp.? Some basic hyperparams (lr, steps) would be appreciated.
+------
+2. The proposed PRM data collection requires model checkpoints from RL baseline training (L252). Therefore, the whole pipeline is: RL baseline training -> PRM data generation collection -> PRM training -> RL training w/ PRM. It makes the first RL baseline training mandatory and increases the computation cost. Did the author explore alternatives, such as using a base policy (e.g., in-house-SFT) to conduct PRM data generation and collection?
+
+### Questions
+1. On the sensitivity to the $\lambda$: L269 mentions that the authors use different $\lambda$ for correct/incorrect code. I'm curious: Did the author try using the same $\lambda$ or other $\lambda$ during hyperparam sweep? Whether the exp. results deviate much from the ones reported in Table 1?
+
+2. Why the x-axis in Figure 4 could have <1 value as the avg. number of collected response per prompt?
+
+3. L366 says the gain is 5.7% on LiveCodeBench and 12.6% on in-house benchmark relative to the RL baseline. What lines do these numbers correspond to in Table 1? I'm guessing I should be comparing the row of Ours-RL (Dense Reward x, Value Int. x) and the last row. But from Table 1 it's 29.8% - 28.2% = 1.6% and 35.8% - 31.8% = 4%.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 3
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+The paper studied using PRM for code generation.
+
+### Strengths
+First attempt to use PRM for code LLMs.
+
+### Weaknesses
+The current paper delivery is poor, many critical questions is not clearly explained.
+
+Weakness 1: 
+The paper discusses using PRM for code generation. But I don't see what specific challenges the proposed method addresses regarding code generation.
+
+The motivation for using PRM in code LLMs (Sec.1 para.2 line 035-043) is that the reward signal is sparse. Sparse reward (or more accurately, the use of bandit feedback) is a common challenge in multi-step reasoning tasks such as math reasoning and code, many research have been conducted to address this challenge, e.g., introducing PRM. 
+- If the paper limits its scope to code generation, the paper should clearly explain what are the specific challenges it address regarding the code generation task. I notice that unit test feedback is mentioned as a cause for the sparse reward, but I don't consider it as a particular challenge specific in the code generation task. In its essence, test cases serve as a means of verification of the holistic response -- same in its role as evaluating the correctness of a holistic response using the gold-answer in math reasoning. The natural question is: Is previous methods to deal with sparse rewards in math reasoning directly applicable to code generation? If not, the current paper does not clearly explain this. 
+- If the paper targets at proposing new methods for PRM, the paper should clearly state what is new about the methods. 
+
+Based on my current understanding of the paper, the paper answers neither of these.
+
+
+Weakness 2:
+For the current version of this paper, I have doubts on the claimed contributions (line 061)
+- From my understanding based on the current paper delivery, the paper only applies previous methods to the code generation task. Reasons detailed as follows:
+	- Most importantly, the paper do not clearly explain what additional challenges arise when collecting process supervision for code generation, in contrast to existing works that collect process supervision in reasoning tasks (e.g., OmegaPRM).
+	- The paper do not clearly explain how the proposed method differ with previous methods (OmegaPRM), except for the application domain (math versus code). 
+
+         OmegaPRM (https://arxiv.org/pdf/2406.06592) introduces the method of automated process supervision using binary search. The paper mentions in line 147: "Instead, we employ an automated approach inspired by techniques used in recent works", but they do not mention how the proposed method differ with OmegaPRM.
+- The paper claims to conduct empirical study of "how to properly and effectively integrate PRMs into RL". I see from section 4.3 that, the paper experiment on using PRM as dense rewards, or/and as Value initialization. Using PRM as dense rewards is studied in MathShepherd (https://arxiv.org/abs/2312.08935). The paper draws a conclusion that using PRM to initialize the value function of PPO does not work, but this paper (https://arxiv.org/abs/2406.03816) explores an effective way to employ PRM for value initialization of PPO. So I don't see valuable contributions in this empirical study. 
+- I question the evaluation of the proposed method. There are no results on common benchmarks for code generations, such as APPS, MBPP, HumanEval
+
+### Questions
+Better articulate the contributions:
+1) What unique challenges arise when applying PRMs to code generation compared to other domains? 
+2) How does the approach specifically address these challenges? What are the specific differences between the proposed methods and previous methods that inspire this method?
+3) The unique challenges in collecting process supervision for code generation versus mathematical reasoning tasks. A point-by-point comparison of their method with OmegaPRM and other relevant approaches, highlighting any novel aspects specific to code generation. 
+4) Analysis of code LLMs trained with the PRM. Any comparisons on how training with/without PRM improve generated code responses. How does the learned PRM generalize? How does PRM help the value estimate during training?
+5) Include results on these specific benchmarks (APPS, MBPP, HumanEval) in their evaluation section. If these benchmarks were not used, provide a clear explanation for why they were omitted and how their chosen benchmarks compare in terms of difficulty and relevance.
+Discuss how their results on the chosen benchmarks might translate to performance on these more standard benchmarks.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 4
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+5
+
+### Summary
+The paper at hand concerns itself with RL fine-tuning of LLMs from unit test feedback, i.e., obtained in code generation tasks by testing the LLM output against a set of unit tests. The main feature here is the use of a process reward model (PRM) to supply dense rewards at multiple points within a generation (token sequence), rather than just at the end (unit test feedback). While PRMs have been used in several domains before, the key contributions are a recipe to obtain a PRM for code as well as ablations on how to best utilize it during RL fine-tuning.
+
+### Strengths
+With direct experience in the paper's domain I found it, for the most part, easy to read and understand. I liked how the paper ablates different parameters for PRM training and usage. The subject of the paper is also of interest within the community as well as for the deployment of code LLMs in products.
+
+### Weaknesses
+My main grievance with this paper is that it has very low reproducibility. Models, training data, and evaluation sets are all proprietary; we don't even learn about the size of the models. It does not seem to be a large model, though, indicated by the "lite" name and by the fact that Table 1 only compares to "mini" models from OpenAI and Google. To add to that, a number of crucial details are missing, e.g., hyper-parameters for SFT and RL.
+
+Throughout the paper, the authors claim that their PRM improves exploration. I don't see this claim verified experimentally. Instead, the experiments show better *generalization* to LiveCodeBench and their in-house benchmark, which is not the same. It would be interesting to see if indeed a larger fraction of training problems is solved during training, or whether more diverse solutions are found.
+
+### Questions
+In Table 1, you compare "Ours-SFT" against "Ours-RL", but the text in L366/367 refers to it as "RL baseline"? Relatedly, since you train the PRM on data produced during the RLHF phase, does that mean "Ours-RL" numbers are the product of two RL training stages? Where is the model after RLHF then (i.e., the "RL baseline")?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2

human_reviews/EvyYFSxdgB.md

@@ -0,0 +1,152 @@
+# DATS: Difficulty-Aware Task Sampler for Meta-Learning Physics-Informed Neural Networks
+
+- Decision: Accept
+- Scores: 6, 6, 6
+
+## Abstract
+Advancements in deep learning have led to the development of physics-informed neural networks (PINNs) for solving partial differential equations (PDEs) without being supervised by PDE solutions. While vanilla PINNs require training one network per PDE configuration, recent works have showed the potential to meta-learn PINNs across a range of PDE configurations. It is however known that PINN training is associated with different levels of difficulty, depending on the underlying PDE configurations or the number of residual sampling points available. Existing meta-learning approaches, however, treat all PINN tasks equally. We address this gap by introducing a novel difficulty-aware task sampler (DATS) for meta-learning of PINNs. We derive an optimal analytical solution to optimize the probability for sampling individual PINN tasks in order to minimize their validation loss across tasks. We further present two alternative strategies to utilize this sampling probability to either adaptively weigh PINN tasks, or dynamically allocate optimal residual points across tasks. We evaluated DATS against uniform and self-paced task-sampling baselines on two representative meta-PINN models, across four benchmark PDEs as well as three different residual point sampling strategies. The results demonstrated that DATS was able to improve the accuracy of meta-learned PINN solutions when reducing performance disparity across PDE configurations, at only a fraction of residual sampling budgets required by its baselines.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper introduces a difficulty-aware task sampler (DATS) for meta-learning of PINNs. The model takes the variance of the difficulty of solving different PDEs into consideration by optimizing the sampling probability of meta-learning. An analytic approximation of the relationship of meta-model and sampling probability is provided to enhance learning. DATS is shown to improve the overall performance of meta-learning PINNs.
+
+### Strengths
+1) Quality: The performance of the approach seems good in the empirical part of the paper and the ablation study is detailed.
+2) Originality: The proposed two strategies to utilize $p^*$ are interesting and the comparison and analysis are comprehensive.
+
+### Weaknesses
+The main weakness is the clarity and correctness in both methodology and experiments
+
+1) Lack of intuitive explanation for the math derivations in section 4.1, making it hard to follow.  For example, the $w_i=  \langle g_{tr}, g_{val} \rangle$ in Eq.9 may be intuitively interpreted as "assign the weight according to gradient similarity between train and valid", I guess.
+
+2) There are some misleading typos and unexplained assumptions in section 4.1.
+* The LHS of Eq(7) should be $l_{\text {val }, \lambda}\left(\theta^{t+1}\right)$, but not $l_{\text {val }, \lambda}\left(\theta^*\right)$ written in Line 5 of this paragraph.
+* And similarly, the LHS of Eq(8) should be $p^{t+1}(\lambda)$ not $p^{*}(\lambda)$.
+* In Line4-5 of this paragraph, the gradient descent of training loss is defined as $\theta^{t+1} = \theta^t-\eta \int_\lambda p(\lambda) \nabla_\theta l_{t r, \lambda}\left(\theta^t\right) d \lambda$, which assumes that training loss is defined as in Eq(11), the so-called DATS-w. But  DATS-rp loss in Eq(12) does not follow this assumption, thus the analysis is invalid for it. 
+* The authors assume that the proposed iterative scheme for $p^*,\theta^*$ converges and that adding regularization further stabilizes the convergence, without explanations. Intuitively, the first-order Taylor expansion is used in Eq(7), thus the step size $\theta$ should be small enough to stabilize it. Additionally, the discrete approximations may also introduce errors.
+
+3) The experimental performance of sampling strategies (Section 5.2 and Appendix C) is not reported clearly. There are only figures in the main text. Fig.3,4,6 are difficult to extract information from since the lines and shades overlap heavily.  And since there are no digital numbers available, it is hard to compare the results. For example, in Fig.C.12, it seems Self-pace has the same performance as DATs.
+
+### Questions
+1) In Fig.3,4,6,  Why do the uniform and self-paced baselines only have higher and lower bounds of errors, not curves at different residual budgets?
+
+2) How does the DATS compare with the Reduced Basis Method, e.g., [1]?
+
+3) Does DATS also perform well on more difficult PDEs, such as with discontinuity and high-dimension? Examples includes the shock tube of compressible Euler's equation and  2d/3d N-S equation.
+
+Refs:
+[1] Chen, Yanlai, and Shawn Koohy. "GPT-PINN: Generative Pre-Trained Physics-Informed Neural Networks toward non-intrusive Meta-learning of parametric PDEs." Finite Elements in Analysis and Design 228 (2024): 104047.
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 2
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper introduces a novel approach to Physics-Informed Neural Networks (PINNs) with a focus on unsupervised learning for parameterized Partial Differential Equations (PDEs). The authors propose optimizing the probability distribution of task samples to minimize error during meta-validation. They transform the problem theoretically into a discretized form suitable for optimization and introduce two optimization strategies: optimizing the residual points sampled and the loss weight across different tasks. The paper presents experiments on several equations, showcasing improvements over baseline methods.
+
+### Strengths
+1. Clarity and Presentation: The paper is well-written, making the novel method and its distinctions from baselines clear. The method is explained in a step-by-step manner, which is easy to follow.
+2. Innovative Approach: The meta-learning method for PINNs is a fresh take on solving parameterized PDEs, and it is well grounded in theory.
+3. Comprehensive Experiments: The authors conduct experiments on various equations, providing a thorough evaluation of their method.
+4. Effective Visualization: The results are presented in a clear manner, with visualizations that aid in understanding the improvements made.
+
+### Weaknesses
+1. Mischaracterization of Data Loss: In Section 3, the paper inaccurately defines data loss as the loss of boundary conditions. This is a mischaracterization, especially for Neumann or Robin boundary conditions, which only penalize normal derivatives rather than resulting in data loss.
+
+2. Formatting and Clarity of Figures: Some figures in the paper could be improved for better clarity and understanding. For instance:
+(1) Figures 3 and 4 would benefit from added grids and key values annotated directly on the figures.
+(2) The scales in Figure 5 are too small to read, making it difficult to interpret the results.
+The authors should review and adjust these figures to enhance clarity.
+
+3. Lack of Comparison with State-of-the-Art: The paper could be strengthened by including a comparison with state-of-the-art methods in the field, providing a clearer context of the method's performance.
+
+4. Limited Discussion on Limitations: The paper does not adequately discuss the limitations of the proposed method, which is crucial for readers to understand the potential challenges and boundaries of the approach.
+
+5. Potential Overfitting: Given the nature of the meta-learning approach, there could be a risk of overfitting to the tasks at hand. The paper could benefit from a discussion on how this risk is mitigated or how the method performs under such circumstances.
+
+### Questions
+1. Clarify Mischaracterizations: The authors should revisit Section 3 to correct the mischaracterization of data loss and provide a more accurate description.
+
+2. Improve Figure Formatting: Enhancements should be made to the figures to improve readability and clarity, as this will aid in better conveying the results and contributions of the paper.
+
+3. Include Comparison with State-of-the-Art: Adding comparisons with leading methods in the field will provide a clearer benchmark of the proposed method's performance.
+
+4. Discuss Limitations and Potential Overfitting: A section discussing the limitations of the method and addressing potential overfitting concerns would add depth to the paper and provide a more balanced view of the approach.
+
+With these enhancements, the paper would offer a more comprehensive and clear presentation of the proposed method, its strengths, and its potential areas for improvement.
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 3
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This paper presents a sampling strategy for meta-learning in physics-informed neural networks (PINNs). The key idea is to conduct sampling based on the difficulty. The authors provide an analytical solution to optimize the sampling probability, with a regularization term. Experiments show improved performance over uniform sampling. An ablation study has been presented to help understand the method. The method also shows better performance under the same budget.
+
+### Strengths
+1. Meta-learning for PINN is a promising solution to generalize PINNs so that we do not have to train from scratch for any new PDE.
+2. The proposed sampling strategy is general and can be combined with the existing meta-learning strategy.
+3. Experiments show the proposed method outperforms the uniform sampling.
+
+### Weaknesses
+1. It is necessary to report training costs in terms of running time. Although the method can improve the sampling efficiency, the sampling strategy itself could be more time-consuming than uniform sampling. It is unclear whether the method can bring actual benefits in training time compared with the baseline.
+2. The method is only tested on three benchmarks. Results on more benchmarks are encouraged to test the generalizability of the proposed method, such as Heat, Wave, and Advection.
+
+### Questions
+What is the actual training time improvement?
+
+### Soundness
+3 good
+
+### Presentation
+2 fair
+
+### Contribution
+3 good

human_reviews/FL1VmOgiO8.md

@@ -0,0 +1,127 @@
+# Sentiment-Enhanced Stock Price Prediction: A Novel Ensemble Model Approach
+
+- Decision: Reject
+- Scores: 3, 1, 3
+
+## Abstract
+Stock price prediction remains a formidable challenge within the realm of financial markets, wherein a multitude of models and methodologies have been under exploration to prognosticate the dynamic behaviour of equities. This research endeavour encompasses an exhaustive examination of extant stock prediction systems, entailing a meticulous assessment of their merits and demerits, concurrently pinpointing discernible lacunae and avenues for enhancement. Subsequently, we harnessed the capabilities of BERT, an exemplar in the domain of natural language processing, to conduct sentiment analysis across a heterogeneous corpus of news articles pertinent to the subject stocks. Additionally, an ancillary sub-experiment was conducted to ascertain the relative impact of three distinct categories of news articles, namely headlines, summaries, and a composite amalgamation of the two, on the efficacy of stock price prediction.
+
+The outcome of this investigative pursuit was the generation of sentiment scores for each trading date, which were subsequently integrated as input features in the training of a neural network. Through a comparative analysis of various neural network models, including but not limited to RNN, LSTM, GAN, and WGAN-GP, we discerned that the WGAN-GP model exhibited the most favourable predictive performance. Building upon these findings, we introduced the FB-GAN model, an ensemble architecture comprising WGAN-GP, which capitalizes on the fusion of historical stock price data and market sentiment scores for enhanced stock price prediction.
+
+Subsequently, a comprehensive evaluation of our approach was undertaken vis-à-vis established models, gauging its performance against five prominent equities, namely Amazon, Apple, Microsoft, Nvidia, and Adobe. In summation, this research makes a compelling case for the integration of BERT-based sentiment analysis within the ambit of stock price prediction. Our initial hypothesis regarding the significant influence of market sentiment on stock price prediction was validated, and our proposed FB-GAN model outperformed all other models. Furthermore, incorporating both the headline and summary of the news article contributed to enhanced stock price prediction compared to utilizing either the headline or summary in isolation.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper studies the incorporation of sentiment analysis into stock prediction. The paper shows the RMSE between the actual stock price and the predicted stock price is lowered when the sentiment score is added on top of the previous stock price.
+
+### Strengths
+The paper presents the idea of leveraging sentiments from news articles for stock price prediction. Every component of the model is explained and the readers can easily understand the structure of the model.
+
+### Weaknesses
+1. The science/engineering contribution of paper is very limited. The paper did not propose new methods of obtaining stock-related news articles, new approaches of de-noising the news articles, new models that are developed to study sentiment-score-augmented stock price prediction etc. The paper gave the reviewer an impression that it is a concatenation of the existing models and data without much deep analysis. 
+
+2. The paper is lack of scientific rigor. 
+* There is no guardrail for the accuracy of the sentiment analysis. That means we don't know the accuracy of the sentiment scores that are fed into the prediction model. 
+* The authors claimed in the model that "a strong correlation between the market sentiments on the stock price movement". But there is no  statistical analysis to back the statement up.
+* We need variable control in the comparison between WGAN-GP and FB-GAN. For example, the model WGAN-GP should have the same input dimension as FB-GAN with the sentiment score set to neural.
+
+3. RMSE is not a good metric overall. It does not translate well to trading strategies because large RMSE does not necessarily mean losing money.
+
+### Questions
+No further questions.
+
+### Soundness
+2 fair
+
+### Presentation
+3 good
+
+### Contribution
+1 poor
+
+---
+
+## Human Reviewer 2
+
+### Rating
+1
+
+### Rating Number
+1
+
+### Confidence
+5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.
+
+### Summary
+The authors tried to enhance the accuracy of stock price prediction via news sentiment.
+
+### Strengths
+N/A
+
+### Weaknesses
+1. Lack of technical contribution.
+2. The overall presentation is in a bad shape.
+3. Weak experiments.
+
+### Questions
+This paper lacks a significant technical contribution, and the overall representation is unclear. I don't think it's a good fit for ICLR.
+
+### Soundness
+1 poor
+
+### Presentation
+1 poor
+
+### Contribution
+1 poor
+
+---
+
+## Human Reviewer 3
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper presents FB-GAN, a stock price prediction model based on sentiment analysis. The authors tested the model on historical price data for 5 stocks, and found FB-GAN to outperform a number of baseline models in average. My main concern about this paper is its novelty. The proposed FB-GAN model appears to be a simple extension of WGAN-GP, and the intuition behind the extra modules is not well justified. In addition, the stock universe is too small, which makes the reported results unreliable. I would encourage the authors to explain in more details the design choice of FB-GAN, and to expand the dataset.
+
+### Strengths
+- This paper presents an extensive study of a number of methods on their performance in stock price prediction.
+- The effect of sentiment analysis to stock price prediction is validated in the paper.
+- The paper is well-written and easy to follow.
+
+### Weaknesses
+- The stock universe in the experiments consists of only 5 stocks, which raises concern about whether the proposed approach would generalize to other stocks.
+- FB-GAN appears to be a naive extension of WGAN-GP. The authors fail to explain the intuition behind the module designs.
+- FB-GAN is worse than WGAN-GP on 60% of the stocks tested.
+
+### Questions
+- Have the authors tested WGAN-GP with sentiment scores as the input features?
+- Have the authors studied the relative improvement after incorporating sentiment scores. An ablation study would let interested readers know the effectiveness of sentiment analysis in stock price prediction.
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+2 fair

human_reviews/HNbNcDsgxL.md

@@ -0,0 +1,237 @@
+# Delta - Contrastive Decoding Mitigates Text Hallucinations in Large Language Models
+
+- Decision: Reject
+- Scores: 3, 1, 3, 3, 5
+
+## Abstract
+Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language processing tasks. Still, they are prone to generating hallucinations—factually incorrect or fabricated content that can undermine their reliability, especially in high-stakes domains such as healthcare and legal advisory. In response to this challenge, we propose Delta, a novel inference-time approach that leverages contrastive decoding to mitigate hallucinations without requiring model retraining or additional training data. Delta works by randomly masking portions of the input prompt, then contrasting the original and masked output distribution generated by the model, effectively mitigating hallucinations through inference-only computations. Delta was evaluated on context-rich QA benchmarks like SQuAD v1.1 and v2, achieving around 3 and 6 percentage points of improvement, respectively. It also showed gains of 7 and 2 percentage points on TriviaQA and Natural Question under-sampling decoding. Delta improved SQuAD v2’s no-answer exact match by over ten percentage points. These findings suggest that Delta is particularly effective when hallucinations arise from contextual ambiguity. Delta presents a computationally efficient and scalable solution for reducing hallucinations in real-world LLM applications by focusing on inference-time enhancements.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+5
+
+### Summary
+This work aims at proposing an inference-time improvement to the regular decoding algorithms by leveraging contrastive-based decoding. Specifically, the authors propose Delta that contrastively decodes from two input contexts, an original context and a masked context with part of the input removed. Experiments in question answering show that the method can help mitigate hallucinations.
+
+### Strengths
+This work tries to tackle an important and interesting problem of reducing regular LLMs' hallucination by modifying the decoding algorithm at inference time. The experiments show good results on SQuAD. The writing is overall clear and structured.
+
+### Weaknesses
+(1) This work misses a reference and comparison to context-aware decoding (Shi et al., 2023). The two methods seem very similar (e.g., Equation (5) in this work is similar to Section 2.2 in Shi et al., where a partial context is removed contrastively). The authors should consider adding a comparison of the two methods' performance and discuss any potential differences between the two. 
+
+(2) This work seems to be experimenting with only one LLM, specifically LLama 3.1 8B. Different model families and different model sizes should be reported. Also, the used Llama 3.1 8B is a non-instructed version. It would be interesting to see experiments with e.g., Llama-3.1-8B-Instruct. 
+
+(3) This work only shows performance improvements in SQuAD. In TriviaQA, Natural Question, CommonsenseQA, and MMLU, the proposed Delta method is worse than the baseline decoding method. 
+
+(4) The mask(z) operation could benefit from exploring more design choices. For example, how would using tokens other than EOS work as the mask token? How would transformations other than masking work (e.g., reordering, summarization/back-translation, randomly copying in irrelevant contexts, etc.)?
+
+References
+Shi et al., 2023. Trusting Your Evidence: Hallucinate Less with Context-aware Decoding. https://arxiv.org/abs/2305.14739
+
+### Questions
+N/A
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+1
+
+---
+
+## Human Reviewer 2
+
+### Rating
+1
+
+### Rating Number
+1
+
+### Confidence
+5
+
+### Summary
+The paper proposes a constructive decoding method to reduce hallucinations, which compares the original logits with the logits computed from corrupt inputs.
+
+### Strengths
+The method is intuitive and easy to implement.
+
+### Weaknesses
+1. The paper lacks plenty of related works and baselines. I suggest the author at least compare the following de-hallucination baselines:
+    + CAD [1]
+    + ITI [2]
+    + DoLa [3]
+    + AD [4]
+    + SEA [5]
+2. The method and equation are similar to CAD [1], where the masking part is the corresponding context for ODQA tasks. However, the paper does not mention and compare this method in SQuAD v1.1/v2, TriviaQA, and Natural Question.
+3. The random masking leads to incompatible with FlashAttention, preventing its practical applications.
+4. The performance improvement of the method is not significant, e.g., a negative improvement on TriviaQA and Natural Question, 0.25 and 0.3 improvements on CommensenseQA and MMLU
+5. The paper is not well written.
+
+[1] Trusting your evidence: Hallucinate less with context-aware decoding
+
+[2] Inference-time intervention: Eliciting truthful answers from a language model
+
+[3] Dola: Decoding by contrasting layers improves factuality in large language models
+
+[4] In-context sharpness as alerts: An inner representation perspective for hallucination mitigation
+
+[5] Spectral Editing of Activations for Large Language Model Alignment
+
+### Questions
+No further questions.
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+---
+
+## Human Reviewer 3
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+This paper introduces using constrastive decoding to mitigate hallucinations in LLM. The so-called Delta method works by randomly masking the input text and adjusting predictions using the masked output distribution. Experiments are conducted on several well-known benchmarks, such as SQuAD, TriviaQA, NQ, CommonsenseQA, MMLU etc. The method brings +4~6 EM scores on SQuAD and no improvements on other benchmarks. The paper is generally well-written and easy to follow.
+
+### Strengths
+Reducing LLM hallucinations is an import research topic and challenging. This paper tries to introduce constrastive decoding (CD) to tackle this problem. The motivation is intuitive and reasonable as CD has been verifyed in other tasks. Different from CD based on Gaussian noise etc. in visual tasks, this paper adopts token-level masking in text LLM. Experiments on SQuAD show the proposed CD method significantly improves the accuracy.
+
+### Weaknesses
+Experiments are conducted on several benchmarks. However, the proposed method only works on SQuAD and shows no improvements on otherss. Therefore, the generalization capability of the proposed method is limited. Why the proposed method does not work well on others tasks is unknown as well.
+
+In addition to visual tasks, CD is also used in other NLP tasks, such as  "Contrastive Decoding Reduces Hallucinations in Large Multilingual Machine Translation Models", "Mitigating Hallucinations and Off-target Machine Translation with Source-Contrastive and Language-Contrastive Decoding" etc..   Whether these existing methods is effective in LLM is unknown. It's better to compare these methods  (or some) with the proposed one to verfiy its effectiveness. 
+
+Ablation studies on technique choices are missing. For example, why chosing masking tokens instead of other variations? what's the best masking strategy?  how the multiply factors in  eq(3) are determined? Why the APC is necessary and how to choose the beta value? etc.
+
+### Questions
+see weaknesses.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 4
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+This paper introduces Delta, an inference-time approach to mitigate hallucinations in large language models (LLMs) without requiring model retraining. The method leverages contrastive decoding by randomly masking portions of input prompts and comparing the output distributions of masked and unmasked inputs. By dynamically adjusting logits during token generation, Delta aims to reduce hallucinated content, particularly in context-driven question-answering tasks. Experimental results on datasets like SQuAD demonstrate significant improvements, with the approach showing particular strength in handling ambiguous contexts and unanswerable questions, achieving up to 14.56% improvement in exact match accuracy for scenarios with no definitive answers.
+
+### Strengths
+- Delta provides an efficient, inference-only solution to reduce hallucinations, avoiding retraining and keeping implementation straightforward.
+
+- The method shows some promising result in improvement question-answering tasks.
+
+### Weaknesses
+- The text sequence masking is currently using random masking. This is a very basic approach. A better approach would be masking important tokens (e.g. entities). 
+- Although Delta performs well in ambiguous question-answering datasets like SQuAD v1.1 and v2, its effectiveness is less pronounced in tasks like TriviaQA and Natural Questions, where factual recall or logical inference is needed. 
+- The exact match and F1 score improvements in CommonsenseQA and MMLU are marginal, which may limit Delta's broader application.
+- Delta is proposed to mitigate hallucination, however, it is evaluated on QA datasets. The method should be extended to other tasks for example, summarization, where hallucination is known to be a problem. 
+- The baseline in the paper is just the vanilla Llama 3 model. Delta should be compared with other baselines with similar approach which is missing in the current work:
+  1. Contrastive decoding using multiple input contexts (including relevant and irrelevant) [1]; 
+  2. Context aware decoding (CAD) [2]
+
+
+
+**Reference**
+
+[1] Zhao et al. Enhancing Contextual Understanding in Large Language Models through Contrastive Decoding
+
+[2] Shi et al. Trusting Your Evidence: Hallucinate Less with Context-aware Decoding
+
+### Questions
+1. Have the authors considered an adaptive masking mechanism where the degree of masking adjusts based on the input’s complexity or ambiguity level? In addition, masking important tokens (e.g. entity) would be a better alternative than random masking. There is a concurrent work with very similar idea [1].
+2. Table 1 and Figure 2 are quite duplicated. Why is Figure 2 needed?
+3. The study uses the Llama 3.1 8B model with 4-bit quantization, which may not be the most competitive model. Had the authors considered conducting further experiments using larger models?
+
+
+
+**Reference**
+
+[1] Gema et al. DECORE: DECODING BY CONTRASTING RETRIEVAL HEADS TO MITIGATE HALLUCINATIONS
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 5
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4
+
+### Summary
+This paper introduces Delta, an inference-time approach designed to reduce hallucinations in LLMs without requiring model retraining or additional data. Delta utilizes contrastive decoding by masking parts of the input prompt, and then comparing the output distributions generated from masked and unmasked inputs. Through this mechanism, Delta aims to detect and mitigate hallucinations effectively. The method was evaluated on multiple benchmark datasets, including SQuAD v1.1 and v2, and showed performance improvements of up to 6% in hallucination-prone scenarios.
+
+### Strengths
+1. The paper is easy to follow.
+2. The method is computationally efficient, making it suitable for integration into existing systems where minimal adjustments are preferable.
+3. The experimental setup on SQuAD and other datasets is well-executed.
+
+### Weaknesses
+1. In this method, the tokens to be masked are selected randomly, which may lead to inconsistent performance.
+2. Delta’s performance declines in tasks that rely more on pre-trained knowledge than contextual cues, such as CommonsenseQA and MMLU. This limits its applicability in domains that require factual correctness without strong context-based cues.
+3. The method's effectiveness appears sensitive to hyperparameter settings, including mask and logit ratios. This limits the method's practicality because it requires setting different hyperparameters for different tasks.
+
+### Questions
+1. Have you evaluated the performance of this method with different hyperparameter combinations on lower-performing benchmarks, such as CommonsenseQA and MMLU?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2

human_reviews/J6QKWjq05Z.md

@@ -0,0 +1,218 @@
+# TreeDQN: Learning to minimize Branch-and-Bound tree
+
+- Decision: Reject
+- Scores: 5, 3, 3, 6
+
+## Abstract
+Combinatorial optimization problems require an exhaustive search to find the optimal solution. A convenient approach to solving combinatorial optimization tasks in the form of Mixed Integer Linear Programs is \emph{Branch-and-Bound}.  Branch-and-Bound solver splits a task into two parts dividing the domain of an integer variable, then it solves them recursively, producing a tree of nested sub-tasks. The efficiency of the solver depends on the \emph{branchning heuristic} used to select a variable for splitting. In the present work, we propose a reinforcement learning method that can efficiently learn the branching heuristic. We view the variable selection task as a tree Markov Decision Process, prove that the Bellman operator adapted for the tree Markov Decision Process is contracting in mean, and propose a modified learning objective for the reinforcement learning agent. Our agent requires less training data and produces smaller trees compared to previous reinforcement learning methods.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This paper introduces TreeDQN, a reinforcement learning algorithm based on DQN for solving Tree MDPs. TreeDQN is trained on the mean squared logarithmic error loss. Specifically, the algorithm is used to learn branching heuristics for branch and bound in the context of mixed integer linear programming problems. 
+
+Empirical results on a set of benchmark problems show some of the advantages of TreeDQN for the purpose of learning a branching heuristic. The results on unseen tasks are somewhat mixed, with some advantage to the branching heuristic learned with TreeDQN.
+
+### Strengths
+The paper presents an algorithm for solving Tree MDPs with the specific application to learning branching heuristics for branch and bound algorithms in the context of solving mixed integer linear programming problems. TreeDQN presents better results on some of the benchmark problems used in the paper.
+
+### Weaknesses
+The presentation is *possibly* the paper's weakest point. The lack of clarity makes me wonder about the value of the value of the contributions of the paper. The main contribution of the paper, TreeDQN, is explained in a single paragraph in the main text. Since the text only states that the algorithms is an adaptation of Double Dueling DQN, I assume TreeDQN is a straightforward adaption of DQN to Tree MDPs.
+
+The paper builds on a couple of previous papers, which I had to skim over in order to understand the present paper. I am not entirely familiar with the line of work of using RL to learn how to branch and I can tell that the paper wasn't written for me. These are the two papers that helped me understand this submission:
+
+Exact Combinatorial Optimization with Graph Convolutional Neural Networks 
+and
+Learning to Branch with Tree MDPs
+
+The example on Mixed Integer Linear Programming isn't very helpful. The tree shown in Figure 1 is uninformative; it simply shows nodes in a tree where the color scheme differs the root of the inner nodes and from some of the leaf nodes. It would have been more helpful to not show a tree and give the reader a full example on how the branch and bound search works. I asked ChatGPT for an example and it gave me an example (without any drawings, of course) that was more helpful than the tree example shown in the paper. 
+
+Overall the background section could be re-written to use less space and pack more information to help the reader understand the work.
+
+I cannot understand the last paragraph of Section 2.2 without reading the paper by Scavuzzo et al. (2022). Here are the question I asked myself while reading that paragraph. 
+
+1. Why do we need to use DFS as node selection or set the global upper bound in the root to the optimal solution cost to guarantee the Markov property? 
+2. The gap between training and testing is due to assuming that one has the optimal solution in training? Why not use DFS and not assume that you have the optimal solution in training? 
+3. How can more efficient heuristics for node selection also induce a gap between training and testing? And why is this important? 
+
+Section 4 lists properties of a successful RL method for this problem, which includes off-policy and "work with tree MDP instead of temporal MDP". Why is it important to learn off-policy? We know of many successful on-policy algorithms for RL, what am I missing here? Why do they have to work with tree MDPs? 
+
+The empirical setting is described in previous papers and the current paper relies on that. How is the training data generated? Do the problems differ in difficulty? Do we have to optimally solve the problem to attain the Markov property to then train the model? If so, how are the problems solved? Assuming that the training instances are easy (one needs to solve them optimally), how does the learned heuristic scale to larger problems? 
+
+The number of seeds also seems to be small (5), for the kind of learning being done. 
+
+Overall, it seems that the paper has some interesting ideas, but I don't fully understand them. The paper was written for people who already knows the details of this line of work, and it isn't friendly to newcomers to the point that the paper isn't self contained.
+
+### Questions
+I would like to hear clarifications on the empirical setup on how the training of the branching function is done, as I listed in the weaknesses section above.
+
+### Soundness
+2 fair
+
+### Presentation
+1 poor
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 2
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+The authors use the TreeMDP framework introduced by Scavuzzo et al. to study RL methods for improved variable selection/branching in branch-and-bound for integer programming with the ultimate goal being smaller search trees. They propose a more stable and sample efficient RL training procedure by choosing a loss function to minimize the geometric mean of tree size during training, and use a deep Q network for training rather than the REINFORCE method used by Scavuzzo et al.
+
+### Strengths
+Branching is a critical aspect of integer programming solvers, and the authors provide an interesting new contribution towards RL based methods for the design of branching rules. The new methods are shown to produce smaller branch-and-bound trees than previous RL based variable selection methods, making this work a promising advance in the “learning to branch” line of work.
+
+### Weaknesses
+Section 2.2 “Tree MDP” needs way more explanation. It more or less assumes familiarity with the Tree MDP work of Scavuzzo et al., and a more self-contained exposition would be very helpful.
+
+The theoretical contribution is very hazy to me. Contraction in mean is not really well-motivated. Does the cited theorem (Jaakkola ‘93) apply to the setting of tree operators here? That seems like a nontrivial assumption that is missing justification. Rather than just including a theorem about contraction in mean, the authors should have a main theorem that states the actual convergence guarantee that follows.
+
+My understanding is that this paper is methodologically very similar to Scavuzzo et al., and only differs in the mechanics of how the RL algorithm is trained. This is discussed in Sections 4.2 and 4.3. In Section 4.2, the main difference is that the authors use a loss function that appears to be selectively picked based on the objective of minimizing the geometric mean of the tree sizes during training/testing. This to me feels like a specific and brittle design choice.
+
+The new method is shown to yield smaller branch-and-bound trees than previous RL based variable selection policies, but no comparison is made to the default settings of any state-of-the-art solver (e.g., Gurobi, CPLEX, SCIP). This is an important comparison that should be included.
+
+Overall the presentation did not convince me that this is a sufficiently novel contribution for ICLR. It seems like the authors just slightly tweaked some aspects of the methodology of Scavuzzo et al. It’s great that these modifications work and yield promising experimental results, but I just did not find the current writeup to be a sufficiently original contribution. The writeup itself also needs quite a bit of work to make it a cohesive, readable, and self-contained (the theory is presented in a very ad-hoc manner without formal definitions) contribution.
+
+### Questions
+“In the B&B search trees, the local decisions impact previously opened leaves via fathoming due to global upper-bound pruning. Thus the credit assignment in the B&B is biased upward, which renders the learned policies potentially sub-optimal.” I understand the first sentence, but what does the second sentence mean? What is “credit assignment”, and why is it/what does it mean for it to be biased upward?
+
+See also questions in the “weaknesses” section.
+
+### Soundness
+2 fair
+
+### Presentation
+2 fair
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 3
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This work studies the variable selection problem in the branch-and-bound algorithm from the point of view of Tree-MDPs, which, instead of the “linear” time-axis present in ordinary markov decision processes, models the decision history as a binary tree.
+They show that under mild assumptions tree-MDPs allow for a contractive Bellman operator, justifying a Tree-MDP version of deep q-learning dubbed TreeDQN. Finally, the authors demonstrate their performance against the “strong branching” baseline and other learnt variable selectors on a large set of synthetic instances.
+
+### Strengths
+Inherently, the idea of modelling variable selection as a Tree-MDP is a great idea as it allows the incorporation of the branch-and-bound structure into the decision process. The modification of the loss function to stably regress towards the geometric mean is also clever and might prove useful even outside the learnt variable selection domain. In general, the presentation of the work is clean and easy to read.
+
+### Weaknesses
+1. Perhaps the biggest limitation is the assumption that the upper bound has to be derivable from the current node or known ahead of time. The authors assert that this (as well as more intricate node selection policies) lead to at most a moderate distribution shift, but never demonstrate this effect.
+2. Another concern is regarding the difficulty distribution of instances. Random instance generation has been known to generate significant amounts of trivial instances compared to real-world equivalents. However, this is a limitation of most prior work on learnt variable selection rules as well.
+3. TreeDQN is also more expensive in terms of wall-clock-time than prior work (especially the IL agent), which can be seen in Figure 4. The paper does not make it clear whether this is due to TreeDQN using a different architecture, or TreeDQN simply creating more expensive nodes during branching.
+4. An important missing baseline in their comparisons is out-of-the-box SCIP, acting as an automatic state-of-the-art hand-crafted tradeoff between SB and cheaper heuristics.
+
+
+The paper needs an extensive re-write in terms of argumentation and clarity.
+
+
+Some more points:
+- Abstract: BnB solver[s] split a task…
+- Abstract: …the Bellman operator adapted for the tree MDP is contracting in mean… - initially I did not understand what you mean with that (only at some later point into the paper)
+- Intro: with [the] Branch-and-Bound algorithm (B&B). |[The] B&B algorithm employs…
+- “The variable selection process is the most computationally expensive and crucial for the performance of the whole algorithm” – is there a reference to prove this? If not, omit this sentence
+- Intro: “problematic”  challenging
+- Intro: “single next state [the] agent”
+- Intro: the contribution list at the end of the section looks like a draft and comes out of nothing
+- Sec. 2: where objective… sentence broken
+- Sec. 2: B&B [-algorithm-] builds
+- Sec. 2: explain “relaxed”
+- Sec. 2: Fig. 1 does not bring much to the table. I suggest to explain B&B with Fig. 1 right from the beginning (add primal/dual, relaxation, variables). This does not cost more space but helps to understand B&B
+- Sec. 2.: [A] straight forward strategy
+- Sec. 2.: [The] tree MDP was proposed by … In the tree MDP [the] value…
+- Sec. 2.: The variable selection process … this paragraph is hard to understand
+- Sec. 3.: “Our work improves…” please add some (technical) argument why this is the case
+- Sec. 4.0: this part takes much space and can be omitted imho. Instead focus on explaining the bullet-point list at the end of 4.0 in more detail. Why must a successful RL method should have off-policy as a property? Policy gradient methods are great, and they are on-policy… Here are a lot of arguments that need more justification.
+- Sec. 4.1 [E]quation3, [E]quation 4
+- Sec. 4.1 is not satisfying to me. The section and with an inequality and tells me that the proof follows from the fact that the tree is finite. Please work out this prove in more detail.
+- Sec. 4.2 the loss function [from E]quation 5
+- Sec. 4.3 we use loss function equation 5 – please re-write
+- Fig. 3 put the description into the plots
+
+### Questions
+- Is the method run on CPU or GPU?
+- What is the performance of SCIP with default parameters on these instances (I.e. reliability pseudocost branching)?
+- What is the model architecture (or more importantly: is it the same for all methods)?
+
+### Soundness
+2 fair
+
+### Presentation
+2 fair
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 4
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This paper extended the on-policy learning to branch method introduced by Scavuzzo et al. in 2022 to an off-policy setting by offering a proof of contraction in mean, a modified mean squared logarithmic error, and an adapted Double Dueling DQN scheme.
+
+### Strengths
+1. The evaluation experiments demonstrate a noteworthy improvement compared to previous work and other state-of-the-art approaches.
+    
+2. The modified mean squared logarithmic error proves to be well-suited for long-tailed distributions of BB tree sizes and exhibits superior performance compared to the mean squared error in the ablation study.
+
+### Weaknesses
+My main concerns about this paper are generalization ability, scalability, and some basic assumptions. Please find details in the questions.
+
+### Questions
+1. Regarding the Assumption in Theorem 4.1: The paper assumes that the probability of having left and right children does not depend on the state because the pruning decision depends on the global upper bound instead of the parent node. However, the global upper bound can change dynamically during the search, which might influence the probability. Does this paper use optimal solutions as upper bounds? Could the authors provide further clarification on this assumption?
+    
+2. Exploring Limited Generalization Ability: In comparing the results presented in Table 5 and Table 3, it is observed that TreeDQN appears to exhibit less stability in the context of transfer tasks. Could you please offer insights or explanations regarding this phenomenon?
+    
+3. A Traditional vs. RL-based Variable Selection Perspective: Traditional variable selection methods rely on human-designed criteria, such as pseudocosts. One advantage of these traditional approach is its applicability to various problem types. On the other hand, current RL-based methods require training an optimal policy for each specific problem. Given the noted limitations in generalization ability, RL methods seem to necessitate training on problem instances of a similar size as the target problems. Could you provide any comments or insights on the potential implications of this limitation? (This question is optional, and your input is welcomed purely out of curiosity.)
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good

human_reviews/JEAlXPYSjC.md

@@ -0,0 +1,167 @@
+# Your CLIP Model Might Be Undertrained
+
+- Decision: Reject
+- Scores: 3, 3, 5, 6
+
+## Abstract
+Contrastive Language-Image Pretraining (CLIP) models exhibit good performance on a range of vision tasks. To improve the performance of this class of models even further, several works have proposed to modify the CLIP training procedure. In this work, we show that it is possible to achieve substantial gains using a much simpler strategy. Specifically, existing CLIP models---especially those trained on smaller datasets---tend to be undertrained. Indeed, we show that extending the training procedure according to a simple heuristic can significantly improve the performance of CLIP models.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper reports the observation that existing popular CLIP training recipe is suboptimal on smaller datasets (under training), and demonstrates clear improvement by resuming the training with a higher learning rate. Experimental results on ImageNet show the effectiveness of the proposed modified training recipe.
+
+### Strengths
+The finding of the paper is clear and simple. The experiments seem convincing. It is a meaningful contribution and the community should move away from the current suboptimal CLIP recipe especially on the CC3M data.
+
+### Weaknesses
+1. Why is the proposed observation of under-training only present with the smaller dataset e.g. CC3M but not with the larger ones like LAION-400M? It'd be very helpful if the authors can provide some insights here, because this seems to be the core contribution of the paper. For example, does the noise level in the image-text dataset affect the fitting behavior in some way?
+
+2. Many existing CLIP recipes rely on high-resolution finetuning after the low-resolution pretraining (e.g.ALIGN, CoCa), which is essentially doing the same finetuning with the extra cycle as shown in Figure 1 (right). Would the proposed approach still benefit models that are already trained with high-res finetuning? 
+
+3. For ablation, I think it'd be cleaner to compare with a baseline that is trained on the same number of epochs so that the only difference is the learning schedule (instead of adding additional training epochs). 
+
+4. The paper writing can use some improvement to expand on the introduction, method, analysis, and related work (still plenty of space left).
+
+### Questions
+See weaknesses in order.
+
+### Soundness
+3 good
+
+### Presentation
+1 poor
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 2
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper revisits the training schedule of the CLIP, especially on those trained on smaller-scale datasets, and finds that a simple continuing training with LR rewinding can significantly improve the CLIP baselines. It demonstrates improvements on 6 ImageNet variants with R50/ViT-B-32/ViT-B-16 backbones and even outperforms some approaches that are designed to be data-efficient (e.g. DeCLIP) when trained on CC12M.
+
+### Strengths
+- The finding of the paper is interesting
+- It establishes a stronger baseline for CLIP training, and questions whether we should solely evaluate data efficient approaches on smaller scale, and its transferability to large-scale datasets.
+
+### Weaknesses
+- The finding does not necessarily transfer to models that are trained on large-scale datasets (e.g. 400M).
+- There is not much analysis / theory on why such behavior exists (or not) on different scales of datasets.
+- It would be interesting to see how the proposed schedule helps when we apply to other approaches in Table 7.
+
+### Questions
+See weaknesses.
+
+### Soundness
+2 fair
+
+### Presentation
+3 good
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 3
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.
+
+### Summary
+This paper presents a training technique using which the performance of CLIP like vision language models can be improved without undergoing any change in objective function or model architecture. Authors of this work shows that CLIP like models trained on small scale datasets like CC-3M or CC-12M might be under trained. To improve their performance, simply finetuning a pre-trained CLIP model again with additional few epochs with a restarted learning rate scheduler is enough. The paper additionally shows that this technique is less effective when tried on CLIP models trained on large scale datasets like LAION-400M. The paper shows additional ablations and result comparisons with previous methods to provide a broad perspective.
+
+### Strengths
+1. This work shows that hyper-parameter and tweaking training strategies for large-scale pretraining of vision-language models plays significant roles in determining their performance on downstream tasks after training.
+2. Restarting the LR scheduler with few-additional epochs is a simple and elegant way to improve performance of CLIP models trained on small scale datasets.
+3. The resulting performance of the model is competitive to prior methods that bring additional objective function or model architecture changes to baseline CLIP model.
+3. Most importantly, this study underlines the crucial need of bench-marking at larger scales to truly reflect improvements due on proposed modifications. Otherwise, showing effects on small scale datasets could be sometimes misleading.
+4. The paper conducts fair comparison and additional ablation studies to provide a broad perspective about CLIP training.
+5. Paper is easy to read and well presented.
+
+### Weaknesses
+1. In my view, this paper presents a effective but somewhat hyperparameter training technique which is more of a engineering trick rather than pure novel contribution. In other words, this paper says that one should use a altered version of multi-cycle LR scheduler instead of a single-cyle LR schedule to improve CLIP performance.
+2. There is little or no analysis on why the proposed trick helps improve CLIP performance. This work can be further supplemented by conducting a detailed analysis on the learned embeddings, for example analysis on modality gap [1], t-SNE visualizations etc.
+3. This work shows performance comparison on zero-shot tasks, but it will be great to see how the learned embeddings provide benefits for adaptation tasks like linear probing or its use on downstream tasks which uses CLIP features in their framework.
+4. Although competitive to prior methods trained on CC3M and CC12M, it is unclear how this technique performs when combined on pre-trained models of prior methods. For example, does this technique shows complementary effect when plugged on CyCLIP pretrained model? (using same CyCLIP objective functions even in the proposed technique)
+5. Similar results with-and without this technique on large scale CLIP models like CLIP-LAION400M shows that this trick is only valid for small scale models with less than 50% baseline accuracy.  
+6. There are missing tables in the paper. For example, I cannot see Table 1 anywhere in the paper.
+
+### Questions
+My main concern is that this paper mainly shows a training trick, which poses questions for novelty of this work. Please refer to the weaknesses section for my additional concerns and queries. 
+
+In summary the paper is nice but it lacks solid technical contributions which can be an significant issue.
+
+### Soundness
+2 fair
+
+### Presentation
+3 good
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 4
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.
+
+### Summary
+This paper makes a simple but intriguing observation that vision-language models trained on even a small dataset such as CC12m can improve substantially (10% zero-shot accuracy), by restarting the training of the model on the *same* data. The paper starts by showing this for on CC12m by training for an extra 15 epochs (Figure 1), then it shows that similar improvements are reached with 2 other models for various number of extra epochs (Figure 3), then shows that the improvements are consistently significant on various OOD datasets (Table 2), then they confirm that restarting training even from early checkpoints can result in faster training (Figure 4), then they extend the idea by repeating the restart by revisiting the idea of cyclical LR schedules and show that a cyclical LR can reach a higher accuracy than a single cycle cosine LR schedule (Figure 5). Finally, they use a cyclical LR schedule to train on the larger LAION400M dataset and conclude that “CLIP models trained on large datasets are less likely to be undertrained.”
+
+### Strengths
+- The results are surprising and promising for future directions on alternative LR schedules for multi-modal training.
+- The paper is easy to read and asks natural questions sequentially that engages the reader. Although the reader is left with many unanswered questions by the end!
+
+### Weaknesses
+- While the paper makes interesting observations, the paper is missing a lot of discussion and potential for extending the observations given the unused page limit. A few unanswered, immediate and simple questions: How many cycles are optimal? How long should each cycle be? Is there a relation between the number of samples and the length of the LR cycles? Figure 5 only presents results with one schedule.
+- Results in Figure 4 and 5 are limited to only one architecture and are not shown to hold for other architectures. Would these results hold for ViT-L/14?
+- The final conclusion is that “CLIP models trained on large datasets are less likely to be undertrained.”. This is based on only one LR schedule and one model that does not provide definitive evidence for the conclusion.
+
+### Questions
+- Where are Figure 2 and Table 1?
+
+### Soundness
+4 excellent
+
+### Presentation
+3 good
+
+### Contribution
+3 good

human_reviews/LBsr2llHz0.md

@@ -0,0 +1,244 @@
+# More Space Is All You Need: Revisiting  Molecular Representation Learning
+
+- Decision: Reject
+- Scores: 5, 5, 5, 5
+
+## Abstract
+Molecular representation learning (MRL) has become pivotal in leveraging limited supervised data for applications such as drug discovery and material design. While early MRL methods relied on 1D sequences and 2D graphs, recent advancements have incorporated 3D conformational information, focusing predominantly on atomic interactions within 3D space. However, we argue that the space beyond atoms is also crucial for MRL, which is overlooked by prior models. To address this, we propose a novel transformer-based framework, dubbed SpaceFormer, which incorporates additional 3D space beyond atoms to enhance molecular representation ability. 
+SpaceFormer introduces three key components: (1) Precision-Preserved Gridding, which discretizes continuous 3D space into grid cells while preserving precision; (2) Grid Sampling, which employs an importance sampling strategy to improve efficiency; and (3) Linear-Complexity 3D Positional Encoding, which extends Rotary Positional Encoding to 3D space to capture pairwise directions and utilizes random Fourier features to efficiently encode pairwise distances. Extensive experiments show that SpaceFormer significantly outperforms previous 3D MRL models across various tasks, validating the benefit of leveraging the additional 3D space beyond atoms in MRL models.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4
+
+### Summary
+The paper introduces a novel architecture for point-cloud molecular data based on the PCA normalization followed by a space transformer. The results seems to be on the SOTA level. However, the Reproducibility Statement is missing and the code is not attached. It's therefore very hard to evaluate the model.
+Technically, the innovations are controversial. A lot of stress in the paper is put on SE3 invariance with respect to global rotation/translation. However, after the PCA, the input representation is already invariant. Another stress is put into absolute positional encoding in SE3. Here, no prior work on geometric attention/transformers is cited, where the same problem has been already solved for point cloud and graph representations. Also, the proposed solution is not rigourosly correct. Finally, I question the gridding technique, why it is needed here as the main blocks operate on continuous positions? Again, sampling additional points in geometric/graph attention and transformer architectures have been well studied in the literature, and must be properly cited.
+
+### Strengths
+The paper is well written and presents a model that outperforms the state of the art. The author even retrained and experimented with the SOTA architecture Uni-Mol.
+
+### Weaknesses
+The Effective Cuboid for Gridding -- the PCA vectors have arbitrary sign, and also uncertainty in the directions if two or more eigenvalues are close in value. As a result, you may have instability, at least with respect to the sign of the frame. This issue has to be at least discussed. Ideally, more experiments need to be run with sign flip of the PCA vectors.
+
+3D Angular Positional Encoding with RoPE is very misleading. You seem to operate on directions from an arbitrarily chosen coordinate center, and measure relative angles with respect to this center. Does it make any sense? It wouldn't make sense even in 2D... Could you please draw the 2D geometry and explain what will be this angle?
+The relative angle is indeed invariant to an in-plane rotation. However, it will change with a different choice of rotation plane, and it will also change with a different choice of the coordinate system (the center of rotation).
+
+The Gaussian kernel approximation has to be validated and explained better, it's very nontrivial why a non-periodic function can be decomposed through products of periodic functions. Even the original paper states "For the Gaussian kernel, k(δ) is not convex, so k is not everywhere positive and δk(δ) is not a probability distribution, so this procedure does not yield a random map for the Gaussian."
+
+More generally, your representation is rather similar to the point-cloud representation with some additional sampling points (which have been studied in the literature), as you are using continuous coordinates as input and full-length attention. Could you please clearly identify the benefits of additionally encoding grid-related parameters?
+
+Sampling empty grids is similar to constructing virtual or aggregation nodes in point-cloud or graph representations. It will be useful to discuss it. Another analogy is stochastic sampling strategy to compute attention better than in N^2, which is used in some recent architectures, please discuss it too.
+
+### Questions
+We incorporate FlashAttention (Dao et al., 2022), reducing the memory cost from O(n2) to linear complexity - this statement is not fully correct. The FalshAttention paper states : We analyze the IO complexity [1] of FlashAttention, proving that it requires 𝑂(𝑁^2𝑑^2 𝑀^-1) HBM accesses where 𝑑 is the head dimension and 𝑀 is the size of SRAM.
+So, it will only scale linearly when the cache size is bigger than Nd (or, formally, is a function of (ND). Which is not the case if the size of your system grows bigger.
+The linear scaling is questionable even looking at the experiments (table 6). Models no 13-14 have an increase in the number of cells by 15%, but the time doubles -- this probably indicates the regime when you saturate the SRAM. Please correct the statements.
+
+"Second, in 3D MRL models, SE(3)- invariant positional encoding" - you already claim to work in invariant frames (after the PCA), the SE3 invariance is not needed here it seems. Motivate the SE3 invariance better please. Maybe you simply gain in performance because of the PCA normalization?
+
+"While the above RoPE-based encoding ... it may be unstable when the entire system undergoes global rotations.", well, it is not supposed to be unstable by design, this was the main idea of RoPE. If it is unstable, please clearly explain why. More generally, you are already using the pose normalization method (PCA), you are INVARIANT to global rotations and translations. So, I believe, this argument is not valid.
+
+"utilizing a 3D positional encoding that is invariant to global rotation and translation" - again, you are in the PCA frames, why do you discuss it?.
+
+In 3D, the actual rotation depends on the sequence of 3 individual rotations (see Euler angle conventions, for example).. Which one do you use? How do you decompose a 3D rotation into 3 rotations about different axes? Does it make any difference? How stable the result will be if you remove the PCA normalization?
+
+Some minor points:
+"which is inspired from microscopic physics where regions near atoms are more crucial." - more crucial for what?
+"regions close to atoms exhibit higher electron density" - better say close to atom centers, or to nuclei centers
+"in computational simulations, coarse-graining is commonly applied to regions farther from atoms to reduce computational cost." - please explain it better and more rigorously
+"Therefore, our sampling strategy is based on the distance from the nearest atom cells." - you can motivate it in a simpler way.
+
+Can the learned representations (Figure 2) be connected to the way you sample empty space using the importance sampling? What will be the representations using uniform or other types of samplings?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 2
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4
+
+### Summary
+I appreciate the opportunity to review this manuscript on "More Space Is All You Need: Revisiting Molecular Representation Learning."
+The paper presents SpaceFormer, a transformer-based framework for molecular representation learning that aims to demonstrate how leveraging 3D space beyond atoms can enhance molecular representations. The work introduces grid-based sampling and combines RoPE and RFF techniques.
+
+### Strengths
+- Novel combination of RoPE and RFF techniques
+- Innovative grid-based sampling approach for spatial representation
+- Interesting comparison with Uni-Mol using randomly sampled empty points
+- Technical innovation in spatial representation methodology
+
+### Weaknesses
+Reproducibility Concerns
+A critical issue that must be addressed immediately is the lack of code availability. This significantly hampers reproducibility - a cornerstone of scientific research. The code should be considered an extension of the experimental setup, and its absence makes proper evaluation challenging. Without access to the implementation details and the ability to reproduce the results, it becomes difficult to fully validate the claims and insights presented in the paper.
+
+Fundamental Research Question and Historical Context
+A significant concern lies in the paper's central research question: "Will leveraging the 3D space beyond atoms improve molecular representation learning?" This framing overlooks substantial existing work in the field. Models like ANI and SchNet have already definitively demonstrated the value of considering complete 3D space through their use of Behler-Parrinello symmetry functions and sophisticated spatial representations. These established approaches already:
+- Capture the complete local chemical environment
+- Incorporate radial and angular features describing inter-atomic spaces
+- Map the entire local 3D space around atoms into feature vectors
+- Consider continuous space rather than just discrete atomic positions
+Therefore, the more appropriate research question should focus on finding novel, potentially more efficient ways to represent and process 3D spatial information, rather than questioning whether such spatial consideration is beneficial.
+
+Technical Innovation and Comparative Analysis
+While the paper's technical approach using grid-based sampling shows innovation, the comparative analysis raises concerns. The baseline model selection appears inadequate, particularly given the availability of established models that already leverage 3D space effectively. The manuscript's results show that SchNet frequently outperforms the proposed SpaceFormer model on the QM9 dataset. The absence of comparisons with models like ANI2 or SchNet requires justification, especially given their relevant capabilities and established performance.
+
+Technical Implementation Concerns
+The use of PCA for coordinate system alignment presents potential issues that need addressing. PCA axes can flip when eigenvalues are similar, potentially leading to arbitrary axis assignments. This could be particularly problematic for molecules with different conformations. The manuscript should address:
+- Whether this possibility was investigated
+- The proportion of affected molecules in the dataset
+- Potential solutions, such as augmentation approaches involving random rotation and translation of voxel boxes
+
+Visualization and Explainability
+Two significant improvements would enhance the manuscript:
+- The latent space structure after pre-training needs thorough examination through ablation studies. Understanding what the model actually learns would improve transparency and help readers better grasp the embeddings' potential applications in downstream tasks.
+- Figure 1's representation of the gridding and grid sampling approach should more clearly demonstrate its 3D nature to avoid ambiguity.
+
+### Questions
+1. How do you address PCA axis flips when eigenvalues are similar?
+2. Why were comparisons with ANI2 and other established models omitted?
+3. Can you provide ablation studies examining the latent space structure (what has been learned)?
+4. How will you ensure reproducibility through code availability?
+5. What percentage of molecules in your dataset are affected by PCA alignment issues?
+6. How do you justify the model's performance given e.g. SchNet's superior results on QM9?
+
+(For more details look at the strengths and weaknesses above)
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 3
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4
+
+### Summary
+This paper presents SpaceFormer model for molecular representation learning which is based on transformer. Compared to previous methods which focus solely on atom/graph level information, SpaceFormer aims to capture spatial information more precisely by taking into account the 3D space beyond just atomic positions. For this purpose, the paper introduces methods for discretizing 3D space for computational efficiency while preserving precise atomic information and benchmarks the model on multiple downstream property prediction tasks.
+
+### Strengths
+1. The paper is well-motivated with physical principles and to my knowledge, introduces a novel MRL framework to explicitly capture the 3D spatial information beyond atomic positions. This allows the model to capture subtle spatial interactions between atoms that maybe crucial for molecular property. 
+2. To handle the infinite nature of 3D space and the computation costs associated with processing grid based representations of molecules, the authors propose extensions of several techniques known in different domains. For example,  the discretization procedure, the 3D angular positional encoding with RoPE and 3D radial positional encoding with RFF are sensible extensions of known techniques for the problem at hand.
+3. The model is tested on multiple downstream tasks, demonstrating SpaceFormer’s effectiveness across a range of MRL applications.
+4. The paper is well-written and easy to read.
+
+### Weaknesses
+1. Although grid discretization captures spatial information efficiently, it does inherently divide continuous 3D space into discrete cells. Could this lead to a loss of spatial continuity, especially when atoms or interactions lie near cell boundaries?
+2. Can the proposed model handle irregular molecular shapes? The experiments considered in the paper are on simpler molecules. When extending it to irregularly shaped molecules, I wonder if the cuboid might include substantial regions without atoms, which could reduce efficiency?
+3. The approach relies heavily on accurate atomic positions and bond distances, which might not always be readily available or precise for all molecular structures, especially in experimental or noisy datasets. Have the authors studied this aspect?
+4. Do I understand it correctly that the proposed method uses a fixed grid size within the effective cuboid, which may not adapt well to regions with varying atomic densities. For densely packed regions, the chosen grid size might be too coarse, while for sparsely packed regions, it may be overly fine? 
+5. The paper studies the model performance under two kinds of data splits: In-Distribution Split, where the sets are randomly partitioned, and Out-of-Distribution Split, where the sets are divided based on fingerprint similarity. The Out-of-Distribution Split is somewhat realistic setting however in real-world drug discovery applications, much more common splitting strategies are scaffold split or temporal split. I would be curious to see how the model performs under such settings.
+6. The paper mentions several related works but the baselines compared against SpaceFormer are rather limited. How does the proposed model compare against works such as NoisyNodes, GraphMVP, 3D-InfoMax etc.?
+
+### Questions
+I have listed my questions in the weaknesses section already. To summarize:
+1. Could the discretization process lead to a loss of spatial continuity, particularly when atoms or interactions lie near cell boundaries? How does the model account for this, if at all?
+2. Can the proposed model efficiently handle irregularly shaped or large molecules? 
+3. Given that the model relies heavily on accurate atomic positions and bond distances, how does it perform on experimental or noisy datasets where such information may be less precise or unavailable?
+4. Does the use of a fixed grid size within the effective cuboid limit the model’s adaptability to regions of varying atomic density? I may have misunderstood this aspect.
+5. Could the authors comment on how SpaceFormer might perform under more real-world splitting strategies (temporal and scaffold splits) for training and evaluation?
+6. Could the authors provide comparisons to how the model performs relative to other recent works, such as NoisyNodes, GraphMVP, or 3D-InfoMax?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 4
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4
+
+### Summary
+This paper proposes SpaceFormer, a transformer-based model for Molecular Representation Learning (MRL) that incorporates information from the 3D space beyond atomic positions. The authors argue that this empty space contains valuable information relevant to molecular properties. SpaceFormer receives as input the discretized 3D space around a molecule as a grid of cells, where cells containing atoms are distinguished from non-atom cells. For better computational efficiency, the non-atom cells are subsampled using an importance sampling strategy. The authors conduct experiments to assess the performance of SpaceFormer on several benchmarks for molecular property prediction.
+
+### Strengths
++ The authors perform ablation studies to analyze the contribution of each component of SpaceFormer, providing insights into the importance of each design choice. The ablation studies on grid sampling are particularly informative, showcasing the trade-off between performance and efficiency.
+
++ The authors extend the Uni-Mol baseline by incorporating randomly sampled empty points, allowing for a direct comparison between an existing atom-based models with and without empty space consideration. This strengthens the claim that SpaceFormer's approach of explicitly modeling empty space within a grid structure is more effective.
+
+### Weaknesses
+- The benchmark results for QM9 shown in Table 1 suggest that the performance of SpaceFormer would be SOTA, but this is not the case. Several published atom-based models designed for regression on molecular properties achieve significantly better results (see e.g. http://proceedings.mlr.press/v139/schutt21a/schutt21a.pdf for an already 3 year old model that is better). Further, the authors should explicitly report the units (I simply assumed that the default QM9 units are used), otherwise no meaningful comparisons to published work are possible.
+
+- While the paper reports performance on QM9, this is a quite old and rather simple benchmark (perhaps comparable to MNIST in computer vision) and the field has since moved to more sophisticated benchmarks, e.g. MD17, MD22, or QM7-X. I think additional experiments and comparisons to SOTA models on (at least one of) these datasets would be very insightful.
+
+- The ablation studies and the comparison to an atom-only model is a good start, but I believe they are flawed: Since the atom-only model processes significantly fewer tokens, it has in a sense "less internal FLOPs" compared to models that also process tokens for empty space. It would be more insightful to see results for an experiment where this is controlled for, for example by increasing the width/depth of the atom-only model to achieve "compute-parity". This would allow to see whether the empty space tokens are useful conceptually, or whether they simply increase the amount of compute the model can leverage.
+
+- The resemblance between electron density and the learned representations shown in Figure 2, as well as the conclusions drawn from it (e.g. that this is evidence that SpaceFormer learns meaningful physical relationships) is questionable. It is trivial to produce a surface that resembles electron density simply by using the distance to the atoms. For example, just take the $\sum_{i=1}^{N} \lVert \vec{r} - \vec{r}_i \rVert^{-2} = 1 a_0^{-2}$ isosurface and you will get something that resembles the electron density much more closely than the representations learned by SpaceFormer. (Here, the sum runs over the $N$ atoms in the structure, $\vec{r}_i$ is the position of the $i$-th atom, $\vec{r}$ is the "query point" in 3D space, and $a_0$ means Bohr, i.e., the atomic unit of length.)
+
+- While the authors motivate incorporating empty space by mentioning some physical principles/facts (e.g. that electric fields and electron density extend beyond atoms), the paper lacks a formal theoretical analysis of why and how this information contributes to better molecular representations. A theoretical foundation would significantly strengthen the paper's contributions.
+
+### Questions
+* Can you provide a quantitative comparison between the learned representations and the corresponding electron density maps? As a baseline, something like a distance-based isosurface (see above) would be useful (the used isovalue should of course be optimized, I just put $1 a_0^{-2}$ for simplicity).
+
+* How do you choose the coordinate system in cases where the eigenvectors determined by PCA are degenerate (for example for molecules with $\mathrm{T_d}$ or $\mathrm{O_h}$ symmetry)?
+
+* The authors mention that the encoding of angular directions may be unstable when the entire system undergoes global rotations. I don't see how this can happen if you choose a coordinate system based on PCA eigenvectors (this should give a sort of "canonical orientation"). Can the authors elaborate?
+
+* How does the performance of SpaceFormer vary with different grid resolutions? Have you experimented with adaptive gridding strategies to address potential limitations for larger molecules or systems with varying densities?
+
+* What are the specific challenges you anticipate in extending SpaceFormer to larger systems like proteins, and what potential mitigation strategies are you considering?
+
+**Additional Feedback:**
+
+* Consider adding a discussion about the computational cost of SpaceFormer during training and inference, compared to other 3D MRL models.
+
+* The paper mentions future work on extending SpaceFormer to larger systems like proteins. However, the current grid-based approach could face scalability issues with larger systems due to the cubic growth in the number of cells. Discussing potential strategies for mitigating these challenges would be helpful.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2

human_reviews/LFn7s8yRUF.md

@@ -0,0 +1,200 @@
+# EXPLORING THE IMPACT OF DATA AUGMENTATION ON LOCALIZED PERSONALIZED AI TRAINING WITH LLAMA3 AND LORA
+
+- Decision: Reject
+- Scores: 3, 1, 1, 1
+
+## Abstract
+With the development of personalized AI models, particularly those emulating characters from novels, games, anime, and films, a significant challenge is the scarcity of suitable dialogue data. These works often feature distinctive styles and character dialogues that may not generalize well to everyday conversations. Data augmentation is crucial for enriching these limited datasets, ensuring sufficient data for learning the target character’s tone and linguistic habits. This paper investigates the impact of various data augmentation techniques on personalized AI models in NLP, specifically focusing on models trained using LLaMA3 through Low-Rank Adaptation (LoRA). We employ different data augmentation strategies, including random deletion, synonym replacement, swapping, random insertion, back translation, and paraphrasing. To provide a comprehensive analysis, we apply these techniques across three distinct datasets, each representing different dialogue styles and contexts. By systematically comparing these methods, we demonstrate their influence on model performance and robustness. This study provides valuable insights into the effectiveness of different data augmentation strategies for enhancing the versatility and robustness of personalized AI systems trained with LLaMA3 using LoRA.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+The paper investigates the relevance of data augmentation techniques for the training of personalized dialogue agents. It gives an overview of existing text augmentation methods, of existing open-weights LLMs and of low-rank adaptation techniques. It introduces a custom dataset based on a TV series and a video game, and a task consisting of enacting a specific character within these environments.
+
+### Strengths
+# Originality
+
+The proposed approach consists of low-rank fine-tuning applied on a training dialogue dataset with data augmentation. The approach in itself is not particularly novel, and its presentation would benefit from a more thorough literature review.
+
+The introduced dataset is novel, and its elaboration would benefit from a more detailed presentation.
+
+# Quality
+
+Only one set of experiments has been run, using one available open-weight model and the dataset that the paper introduces. This is insufficient to get to sound conclusions as to which data augmentation techniques are better in general for the task of dialogue modeling.
+
+# Clarity
+
+The paper is easy to follow, but misses key details in how the dataset has been constructed: the paper should mention where the data comes from, how it was processed, what are the key characteristics (# of episodes, of tokens, of characters...), how the training/validation split was done...
+
+# Significance
+
+The understanding of which data augmentation techniques are most relevant in order to adapt LLMs to small datasets is of prime importance for the field. However, it seems that the proposed method and the results do not give very strong indications of winning strategies.
+
+### Weaknesses
+Regarding the proposed dataset:
+* It is not clear where the training data comes from and how it was curated
+* It is not clear how the proposed dataset relates to existing work, and which specific problems it aims at solving
+* There is no measurement of training data leakage. Given their popularity, it seems likely that these TV shows are quite prevalent in LLM training data
+
+Regarding the fine-tuning results:
+* There is no comparison with a baseline that does not use data augmentation
+* There is no comparison to any few-shot prompting scenario
+* It is not clear what the fine-tuning for a specific character consists of: how is the autoregressive loss computed when multiple characters are talking?
+
+Regarding the paper presentation:
+* The paper spends several pages presenting topics that are not directly relevant for the subject at hand. While understanding low-rank fine-tuning is a topic of importance, it is not particularly relevant to spend 1 page on it, at the expense of experimental details and results.
+* The abstract mentions 3 different datasets, but the results cover only two of them.
+
+### Questions
+* Could you check the supplementary material? The content provided as supplementary material is not a zipfile and can't be decompressed by zip utility.
+
+* How well would prompt-based solutions fare in this setup?
+
+* Would the results still hold when used on existing dialogue datasets?
+
+### Soundness
+1
+
+### Presentation
+3
+
+### Contribution
+1
+
+---
+
+## Human Reviewer 2
+
+### Rating
+1
+
+### Rating Number
+1
+
+### Confidence
+5
+
+### Summary
+This paper investigates the effectiveness of various data augmentation techniques for training localized, personalized AI models using LLaMA3, fine-tuned through LoRA. The study employs methods like random deletion, synonym replacement, backtranslation, and paraphrasing to enhance the quality and robustness of personalized dialogue models trained on character-specific data. Experiments are conducted on two datasets with distinct linguistic styles, aiming to provide a detailed comparison of augmentation techniques in low-resource, domain-specific training scenarios.
+
+I find that while this paper tackles a relevant problem and provides an insightful comparison of augmentation techniques, its contributions remain limited. The paper needs a more novel methodological approach or a broader set of experiments that highlight its findings' relevance to a wider range of personalized AI training scenarios. Therefore, I unfortunately cannot recommend acceptance for this paper; however, I have included some suggestions for improvements in my comments below.
+
+### Strengths
+1. **Clear and Relevant Problem Statement**: The scarcity of personalized dialogue data is an issue of practical importance for character-based AI models. The study addresses this with a focus on data augmentation, an area less explored in NLP.
+2. **Multiple Metrics Used**: The paper evaluates on both training and validation loss, as well as BLEU and ROUGE scores, providing valuable insights for optimizing dialogue data augmentation.
+3. **Resource-Conscious Approach**: The authors' choice of LoRA for fine-tuning and Unsloth for training efficiency is thoughtful and appropriate for low-resource environments, demonstrating an understanding of practical constraints in the field.
+
+### Weaknesses
+1. **Incremental Contributions**: While the topic of data augmentation in NLP is valuable, the paper’s contributions appear largely incremental. Many of the techniques explored (e.g., synonym replacement, backtranslation) are well-known and widely used. The paper could benefit from exploring more novel augmentation techniques tailored to character-based dialogue data's nuances. Additionally, *about half of the paper is describing past techniques* like EDA, Paraphrasing, LoRA, etc (specifically, Section 2, and 3), and there is nearly not enough content for this paper to be considered an impactful contribution (figure 3 has too little information to be taking a whole page and Table 1 font seems overtly big as if just to fill some space). 
+2. **Limited Dataset Scope**: Only two datasets are used, both of which are highly specific in context (e.g., Paimon from *Genshin Impact*). This narrow focus raises questions about the generalizability of the findings to other personalized dialogue systems or broader NLP applications. Including a wider variety of datasets or character types would add robustness to the conclusions.
+3. **Paraphrasing Overfitting Concerns**: The paper notes overfitting issues with the paraphrasing technique but does not fully explore ways to mitigate this. Although the authors acknowledge the limitations of paraphrasing with large models in domain-specific contexts, a more thorough investigation into alternative paraphrasing techniques or domain adaptation methods would strengthen the study.
+4. **Issues With the References**: a) There are some invalid entries in the references, such as, "Unknown Author" and b) The reference list is too short. I suggest the authors to conduct a more thorough review of the current literature.
+
+### Suggestions
+
+1. **Broaden Dataset Variety**: I think the authors are already aware of this problem, but including datasets that represent a wider range of conversational styles or character personalities would improve the generalizability of the results. Alternatively, conducting a detailed analysis of *why* certain augmentation techniques are particularly suited or unsuited to the specific characteristics of these datasets would be more interesting.
+2. **Investigate Advanced Augmentation Techniques**: Consider exploring novel or advanced data augmentation techniques that could better align with the semantic and contextual requirements of character dialogue, such as sentiment-aware or context-sensitive paraphrasing. Additionally, using domain-adaptive techniques for paraphrasing may mitigate overfitting while still generating diverse examples. There is also a huge body of work on using off-the-shelf LLMs for data augmentation that the authors should explore.
+3. **Refine Analysis of Computational Constraints**: Given the choice to work with the smaller LLaMA3-8B model, it would be beneficial to address the specific trade-offs involved more explicitly. This discussion could include any adjustments made to balance memory efficiency with model performance and how this may limit model generalization capabilities.
+4. **Improve Paper Organization**: We don't need to describe past methods like EDA, LoRA, etc. in so much detail. Instead, I would suggest focusing on including more strong baselines (like LLM-based data augmentation).
+
+### Questions
+N/A
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 3
+
+### Rating
+1
+
+### Rating Number
+1
+
+### Confidence
+5
+
+### Summary
+The paper explores different data augmentation strategies (i.e., random deletion, synonym replacement, swapping, random insertion, back translation, and paraphrasing) to train personalized AI models. The model is evaluate by using LoRA and two dataset proposed by the authors.
+
+### Strengths
+None
+
+### Weaknesses
+- Overall lack of clarity. It is unclear to me what is the major claim of the paper. The paper mention several techniques on data augmentation methods, but none of the it's explained in details nor formally. The results are also mysterious, leaving out most of the details. 
+- The results on the paper are non-sensical, the plots are meaningless, showing no particular trends. 
+- The paper contains screenshots from other papers (llama3), without a clear citations.
+
+### Questions
+None
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+---
+
+## Human Reviewer 4
+
+### Rating
+1
+
+### Rating Number
+1
+
+### Confidence
+5
+
+### Summary
+This paper studies the use of text based data augmentations techniques for learning models which can role play as specific characters in diverse scenarios. The paper compares 6 data augmentations strategies on two characters whose content they sample from games as the basis of training Llama 3 using LoRA.
+
+### Strengths
+- At present, the paper does not offer any clear value to the community. The problem of data gathering and evaluating LLMs ability to roleplay is an important challenge, but the present work fails to engage with the broader literature both on these particular models and more specifically in this research space nor to offer carefully controlled experiments which answer questions from the wider field.
+
+### Weaknesses
+- The methods studied don't address the core challenges the authors themselves present. While the authors highlight issues that most training data used for this task occurs in a particular thematic framework as a limitation, the data augmentations they use, which are all methods which have been carefully studied in existing work but are not attributed, only perform local word and structure level permutations on data.
+- The paper, as written now, spends much more time explaining (somewhat tangential) background than seems necessary. The concrete experimental setup doesn't begin until almost halfway through the work.
+\
+\
+For example, the section on Unsloth in section 2.4 and the inclusion of external model evaluations on benchmarks that are unrelated to personalization in *both* Figure and Table 1 have little to do with the core research questions. These sections distract from the narrative of the paper and should be moved to the appendix to make room for more substantive experiments on the primary topic.
+- The work primarily presents a study on just two characters from role-playing that the paper both introduces itself and compares minimally to any external baselines. It is poorly contextualized in the field without evaluations on datasets from the broader community or comparisons to even simple baselines like prompting the model to roleplay.
+- There are many signs that the work was minimally refined before submission. Despite having only 11 citations , in these minimal citations there are obvious errors such as listing "Journal Name" as the publication venue and "Unknown Author et al" as the authors (the work in reference has both publicly known authors and a venue of publication: https://openreview.net/forum?id=rc2h1h89aDi). The figures have mislabeled legends (in figure 3 `BT Loss` is repeated twice) and there are typos where content is duplicated at the start of paragraphs (Line 79: "Wei and Zou (2019) Wei & Zou (2019) introduced")
+
+### Questions
+- Why were the data augmentations techniques not compared to finetuning *without* any data augmentations? This seems like the simplest baseline, so it's unclear why this was not evaluated.
+- These BLEU and ROUGE scores seems extremely low. Are there any examples of model outputs or understanding of why they are so low? As is, these scores are likely to be reached by a model that just outputs any Chinese tokens rather than one that is personalized. 
+- Similarly, the decreasing loss shown could simply be a function of the model learning to always output Chinese, rather than the other languages it can output. How did you evaluate the forgetting effects of your method?
+- In what context are the models evaluated on these new datasets? Is it testing whether the model is able to repeat the phrases a character speaks in it's existing context really a sound way of evaluating whether it is well suited to role-playing for downstream use?
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1

human_reviews/LebzzClHYw.md

@@ -0,0 +1,188 @@
+# Instructive Decoding: Instruction-Tuned Large Language Models are Self-Refiner from Noisy Instructions
+
+- Decision: Accept
+- Scores: 8, 8, 6, 8
+
+## Abstract
+While instruction-tuned language models have demonstrated impressive zero-shot generalization, these models often struggle to generate accurate responses when faced with instructions that fall outside their training set. This paper presents \textit{Instructive Decoding} (ID), a simple yet effective approach that augments the efficacy of instruction-tuned models. Specifically, ID adjusts the logits for next-token prediction in a contrastive manner, utilizing predictions generated from a manipulated version of the original instruction, referred to as a \textit{noisy instruction}. This noisy instruction aims to elicit responses that could diverge from the intended instruction yet remain plausible. We conduct experiments across a spectrum of such noisy instructions, ranging from those that insert semantic noise via random words to others like `opposite' that elicit the deviated responses. 
+Our approach achieves considerable performance gains across various instruction-tuned models and tasks without necessitating any additional parameter updates. Notably, utilizing `opposite' as the noisy instruction in ID, which exhibits the maximum divergence from the original instruction, consistently produces the most significant performance gains across multiple models and tasks.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper introduces an approach for improving the performance of instruction-tuned LLMs on classification and generation tasks. The key idea is to contrast the token generation probabilities conditioned on the true instruction with probabilities conditioned on noised instructions. The paper performs experiments across four model families (Tk, T0, Alpaca, and OpenSNI) on the UnNatInst and SupNatInst datasets, finding that the proposed approach generally produces improvements across tasks and model families over a standard decoding baseline.
+
+### Strengths
+S1) The approach is simple but well-motivated: it doesn't require access to additional models (as Contrastive Decoding does) or access to model internals beyond token probabilities. 
+
+S2) The experimentation was comprehensive, and I appreciated the analysis and ablation experiments, especially the "degradation to enhancement" part, the generalization experiments, and the analysis of sensitivity to the \epsilon parameter.
+
+S3) The results were convincing to me: the approach produces consistent improvements, with substantial benefits in cross-dataset generalization.
+
+S4) The paper was overall very-well written.
+
+### Weaknesses
+W1) While I appreciated the analysis of sensitivity to the \epsilon parameter, given that there is only a relatively small region of \epsilon where the approach outperforms the baseline, I'd appreciate more discussion of how \epsilon was chosen and whether it was the same across all tasks in each experiment.
+
+W2) There's a bit of room to improve the analysis and discussion of it, see questions and comments below. 
+
+W3) This is a minor weakness, but I didn't really understand the connection to "anchoring". Anchoring seems to be about sensitivity to the first thing appearing in the text, while this method doesn't privilege any particular positions in the text but contrasts against the negative prompt at all positions.
+
+W4) While the related work was generally good, there is room to mention a few other related papers (below) on contrastive methods for generation. But I also don't think this is a major weakness, as the application to instruction following is new.
+
+- Li et al. A Diversity-Prompting Objective Function for Neural Conversational Models. NAACL 2016
+- Shen et al. Pragmatically Informative Text Generation. NAACL 2019
+- Ho and Salimans. Classifier-Free Diffusion Guidance. 2021
+
+### Questions
+Q1) How was the \epsilon value chosen for the important experiments in the paper (e.g. Table 1 - Table 3)?
+
+Q2) I was confused by the discussion of the smoothing coefficient in 3.3. It says that the typical choice was 0.3, but performance tends to decline with positive \epsilon values. And the equation for decoding is z - \epsilon \tilde z, which makes it seem like \epsilon should be positive in order to get a contrast.
+
+Q3) Can any intuition be given for why the approach seems to work best in the generalization setting (3.3 "Generalization Capabiltiies")?
+
+Q4) What method is used to choose the random words from NLTK (i.e. what word list is used? something involving antonyms?)
+
+Other comments:
+- The discussion in the "ID with Opposite" section hints that some of the baseline models might be mode-splitting, and that (at least on these classification tasks), approaches might benefit from a decoding approach that uses min-Bayes risk or consensus decoding: group together words based on their similarities and choose to generate a word from a large group.
+- I wasn't really clear on the support for the conclusions in "Visualization of Embeddings", that as models get better at understanding instructions (e.g. are larger), performance of ID increases
+- "aims to steer the models toward particular" in Section 1. Should "toward" be "away from"?
+- "pragmatics" in the motivation for Label Adherence and Label Coherence is a bit misleading; this is much closer to "semantics".
+- it would help to make some of the table/figure captions more self-contained, e.g. defining * in Table 3 caption.
+
+### Soundness
+4 excellent
+
+### Presentation
+4 excellent
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 2
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This paper proposes a simple yet effective method named instructive decoding (ID) to boost the performance of instruction-tuned models. First, an noisy instruction is crafted to elicit undesired response from instruction-tuned models; then ID refines the next-word logits based on the logits from the noisy instruction. The method is training-free and demonstrates improvement in many instruction-following tasks.
+
+### Strengths
+- the method is simple and effective.
+- extensive experiments and analysis
+
+### Weaknesses
+-it’s still unclear (for a practitioner or researcher) to what extent the noisy instruction would be helpful for a new task. Maybe some qualitative analysis in those no-improvement tasks would provide some intuition on why the noisy instruction is harmful in the proposed instructive decoding.
+
+### Questions
+- is it possible to further fine-tune the model via the new logits (i.e.,  $z_t - \tilde z_t$ in Algorithm1)
+- I wonder if the improvement from ID correlates with the response length required for a task. E.g., is ID more useful in tasks that require long response generation?
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 3
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper introduces a method called Instructive Decoding (ID) designed to enhance the performance of instruction-tuned LLMs when they encounter unfamiliar tasks due to the lack of diverse annotated prompts during training. Unlike other methods that rely on expanding the set of instructions, ID manipulates the model by adding 'noisy instructions' that alter the logits of the next-token prediction contrastively. These 'noisy instructions' are essentially counter-instructions that generate contrasting logits, which are then applied to the original logits to refine the prediction distribution.
+
+Evaluations show that models augmented with ID have an edge over basic models in zero-shot generation and better adherence to instructions for certain tasks. The main metric used for evaluation is the zero-shot Rouge-L score on unencountered tasks.
+
+### Strengths
+1) The Instructive Decoding (ID) approach is interesting. It offers an alternative to enhancing LLMs using a broader set of instructions and doesn't necessitate further fine-tuning.
+
+2) The methodology is straightforward and the evaluation metrics, especially those in the table, offer a comprehensive view of the results.
+
+### Weaknesses
+The results, although promising, are mixed. From table 1, only about half of the tasks demonstrated improvements, while the rest experienced a decline in performance.
+
+The paper's rationale needs clarity, especially from a machine learning perspective. It seems there's an assumption that noise-based responses will always produce contrasting predictions at every word level. For instance, if the base prediction is 'xxxx', and the noisy prediction is 'xxxx is wrong', the 'xxxx' part remains consistent, which can be misleading.
+
+The paper references the "anchoring effect" but doesn't provide a clear link between it and the proposed method. The model doesn't directly witness the output from the noisy instruction. Instead, this output is indirectly factored into the next-word prediction logits computation. This differs from the anchoring effect, where individuals receive specific information before arriving at a final judgment.
+
+### Questions
+Can the authors shed light on the underlying rationale behind the proposed method, especially concerning the prediction of logits? How can it be ensured that the noisy responses will always lead to contrasting predictions at every word level?
+
+How is the "anchoring effect" relevant to the proposed method? The paper seems to lack a detailed explanation connecting the two.
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 4
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper proposes a new decoding method, instructive decoding, which generates tokens with logit scores constrasted with the logits under perturbed instructions. Extensive experiments show that such a technique is effective across a wide range of tasks,
+
+### Strengths
+- The presented idea is intuitive and interesting.
+- Extensive experiments and ablations show that the idea works.
+- The paper is generally well-written.
+
+### Weaknesses
+- The empirical gain in Table 1 seems incremental (e.g. only 1-2 points of overall improvement) 
+- The idea is not particularly groundbreaking
+
+### Questions
+- Would there a better way to present the results in Table 1? Currently it feels incremental since the improvement is only 1-2 points, but I could see other arguments against this (e.g., increasing from 3b to 11B also only does not improve much). Would there be a better way to help the readers better contextualize the improvement (or find datasets that are more discriminative between weaker and stronger systems)?
+
+### Soundness
+4 excellent
+
+### Presentation
+4 excellent
+
+### Contribution
+3 good

human_reviews/NZ5KXXDv1T.md

@@ -0,0 +1,177 @@
+# Reinforcement Learning based Image Generation via Visual Consensus Evaluation
+
+- Decision: Reject
+- Scores: 3, 3, 3, 1
+
+## Abstract
+Image generation models are typically trained using the L2 or cross-entropy loss, and evaluated using IS or FID. The inconsistency between the training and evaluation metrics results in suboptimal model performance. To this end, we explore to address the aforementioned issue by finetuning pre-trained generative models with the reinforcement learning. Considering that current evaluation metrics can not be used as training objects since obtaining an accurate score typically demands more than ten thousand images, we introduce an innovative automated metric that captures consensus as a reward signal of the reinforcement learning for finetuning image generation models. It exhibits strong correlation with commonly used metrics such as FID, and demonstrates better robustness to the number of images than FID. Experiments indicate that when introducing varying degrees of noise to the generated images, such as ImageNet contamination or Gaussian noise, our metric quantifies the level of disruption more accurately than IS. By finetuning generative models with our proposed method, we boost the performance for image generation on multiple benchmarks like LSUN 256x256 and ImageNet 64x64.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper has two contributions. 1) It introduces CIGE as a new metric, which can be used both on generative model training and evaluation. 2) It proposes a new RL-based loss term to train the consistency model. Experimental results on LSUM bedroom and Imagenet datasets show that the proposed method has better performance than the compared baselines.
+
+### Strengths
+- [method] CIGE shows a high correlation with the FID while being robust to fewer images, which is very interesting.
+- [experiment] Experimental results on two datasets show that the proposed method has better performance than the compared baselines.
+
+### Weaknesses
+- [writing] This paper seems to be a rushed work and contains a lot of flaws, making it not ready for publication. Here are a few flaws I discovered: TF-IDF misses the reference, Table 3 has a wrong caption, FID in Table 1/2 is inconsistent with Table 3, Fig. 4 is placed after the reference, Fig. 4/5 are blurry, and Table 4/5 are not correctly rendered. 
+- [experiment] Although I like the idea of this work, especially using CIGE as a new metric on par with FID, Table 3 shows that using RL only has a marginal improvement compared with the baseline model without RL. In the LSUN bedroom dataset, using RL even makes the result worse. The above results cannot support the main claim and the effectiveness of the proposed method.
+
+### Questions
+No
+
+### Soundness
+2 fair
+
+### Presentation
+1 poor
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 2
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper present a new metric for image generative models to further improve the trained model to generate images with better quality.
+
+### Strengths
+1. The paper design a new training metric that is closer to the evaluation metric and then finetune the models with RL.
+
+### Weaknesses
+1. The paper is poor-written, hard to follow, and there are many formatting issues
+ - like table 4 and table 5 referenced in sec. 4.2.2.
+ - Or what is CIGE? It does not sound like Consensus-based evaluation metric for Image Generation, this should be CEIG?
+ - Some reference are missed in sec. 3.1.  
+ - in sec. 4.1.1, it directly refers to fig. 5 in the appendix, if this figure is important, it should not be in appendix. 
+ - Table 1, what is NFE? for FID, lower is better, why the OURS one is bold? and which model it fineuted? 
+ - In sec 4.2, it mentioned IS is one of metrics but there is no IS in Table 1 and 2.
+- many typos, e.g. section 5, "We also e introduce", what is the e for?
+
+### Questions
+1. If the proposed CIGE follows the behavior of CLIPScore, why not just use CLIPSore?
+2. Is there any figure related to the description in sec. 4.1.2? Figure 1? 
+3. What is cogview model in 4.1.2? what is the reference?
+4. For Table 3, what is number of the proposed metrics? And the difference between w/ and w/o are very small, why is that?
+
+### Soundness
+2 fair
+
+### Presentation
+1 poor
+
+### Contribution
+1 poor
+
+---
+
+## Human Reviewer 3
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+The paper proposes a novel perceptual score following the paradigm of CIDEr for text. Thereby, a generated and a reference image are embedded using a VQ-VAE image tokenizer and the embedding sequences are treated similar as language tokens in CIDEr: generated and reference n-grams are counted and compared. In order to use the score as a loss in training generative models, the authors rely on reinforcement learning. The paper studies how the proposed score behaves under synthetic perturbations including additive noise and image distortions. The score is further applied to improve the quality of diffusion/consistency models on LSUN (256px) and ImageNet (64px).
+
+### Strengths
+To my knowledge the idea of using a concept similar to CIDEr to create a visual quality score is novel, and I find it pretty inspiring. Further, there seems no metric/score available with similar characteristics to FID, but which can be used as a loss (except maybe a GAN discriminator). In that sense the paper aims to tackle a highly relevant research problem.
+
+### Weaknesses
+Unfortunately, the paper makes a rather unpolished impression (see minor issues below), but also has some larger issues.
+
+- Most importantly, the metric relies on a reference image for every generated image. This makes it unsuitable for unconditional generation, so it is unclear how the metric is applied to the LSUN bedroom data set (please clarify). For class-conditional ImagNet, one could use the images per class, but I’m not sure if the metric could handle the necessary diversity. Either way, details are missing on how the score is computed for the different data sets.
+- According to Section 3.1 each image is mapped to 2x2 tokens. This sounds like a very low number, and the VQ-VAE would need to be extremely compressive to map a 256px image. I doubt it could meaningfully capture a lot of visual detail necessary for a good assessment of visual quality.
+- According to Figure 1, the proposed “CIGE exhibits a strong correlation with FID”. This is clearly not true if the legend of that figure is correct. Further, I would expect a good visual score to capture the visual difference between FID of 100 and 30, which is quite substantial for a human.
+- It seems the proposed method is not effective, despite claimed in the text: In Table 3 the FID and Inception score are almost identical with and without the proposed method/RL.
+
+
+Minor issues:
+- There are quite a few typos: For the first page: “training objects” -> “training objectives”, “lossand” -> “loss and”, “particuler” -> “particular” 
+- Notation collision for symbols in Section 3.1: g used for generated image, quantity in Eq. (1); n used for sequence length and as a superscript/exponent.
+- Inception score is repeatedly computed on LSUN Bedrooms, even though it is not well suited for data other than imagenet.
+- Discussion and references on Precision/Recall (Sajjadi et al. 2018) is missing from related work.
+
+### Questions
+Please see weaknesses for points that could be addressed in the rebuttal.
+
+### Soundness
+2 fair
+
+### Presentation
+2 fair
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 4
+
+### Rating
+1
+
+### Rating Number
+1
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper addresses the issue of inconsistency between training and evaluation metrics in image generation models. It proposes a method that fine-tunes pre-trained generative models using reinforcement learning, with a novel automated metric that measures consensus as a reward signal. This metric correlates strongly with commonly used metrics like FID and demonstrates greater robustness to variations in the number of images. Experiments show that by applying this method, the performance of image generation models improves across various benchmarks.
+
+### Strengths
+The paper tries to proposes a metric that can match the human consensus of model quality with fewer number of sample unlike FID. However, the contribution is hindered by poor presentation quality.
+
+### Weaknesses
+1. The presentation and writing is pretty bad. For example, 
+
+- Space is not correctly used
+- Some terms like abbreviation CIGE and TF-IDF come without explanation first. Better to give a discussion about TF-IDF and how it is related to your approach.
+- Symbols are messy.
+
+2. Experiments are not convincing. For example, in Table 3, how does the number support your approach?
+
+3. Many relevant works that tried to address the issue of FID have been proposed, and they should be included in the discussion also.
+
+Thus, the readability is pretty bad. I would suggest the authors rewriting the paper and submitting to another venue.
+
+### Questions
+See above.
+
+### Soundness
+2 fair
+
+### Presentation
+1 poor
+
+### Contribution
+2 fair

human_reviews/OcaKeyGb0K.md

@@ -0,0 +1,199 @@
+# A unified theory of scene representation learning and object representation learning
+
+- Decision: Reject
+- Scores: 3, 3, 3, 3
+
+## Abstract
+The goal of representation learning is the unsupervised learning of simple and useful representations that model sensory input.
+Various methods have been proposed in representation learning, but a unified theory has not yet been established.
+Two problems exist in the representation learning of a visual scene that contains multiple objects: scene representation learning and object representation learning.
+Scene representation refers to decomposing a single visual scene that contains multiple objects into a combination of multiple individual objects.
+Object representation refers to decomposing a single object into a combination of multiple attributes, such as position and shape.
+Scene representation learning and object representation learning have been formulated in different ways in previous studies.
+Recently, Ohmura et al. (2023) proposed a theory of object representation learning in which transformations between two objects are learned to satisfy algebraic independence so that one attribute of a single object can be transformed while the other remains invariant.
+In existing methods of object representation learning, independence is often imposed between scalar variables, whereas theory based on algebraic independence successfully weakens the constraint from between scalar variables to between latent vectors.
+The latent vector is also used to represent an individual object in existing methods of scene representation learning because such a vector can contain more information than the scalar variable.
+Furthermore, one of the main components of algebraic independence is commutativity.
+Existing methods of scene representation learning typically represent a visual scene as the sum of multiple object representations, and the sum satisfies commutativity.
+We focused on the commonalities between object representation learning and scene representation learning: constraints between latent vectors and commutativity.
+We proposed a unified theory based on algebraic independence that explains both scene representation learning and object representation learning.
+We validated our theory in experiments on an image dataset that contained multiple objects.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+A recently proposed theory for object representation learning is used to develop a unifies theory of object and scene representation learning.
+Experimental results using two simple datasets are presented.
+
+### Strengths
+Extension of a recent ":theory "of object representation learning to integrate scene representation learning and object representation learning.
+
+### Weaknesses
+The proposed theory should be validated using more challenging datasets. The proposed theory does not account for background. The modeled transformations are too simplistic. How will this work when one object is occluded by another or when 3D translational and rotational transformations are considered?
+
+### Questions
+Generalize the proposed theory to more challenging datasets and transformations.
+
+### Soundness
+2 fair
+
+### Presentation
+2 fair
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 2
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.
+
+### Summary
+This paper proposes a unified theory approach for learning object and scene representations. The authors draw inspiration from Ohmura et al. and propose an algebraic approach to object-scene representation learning. The authors claim to have unified the learning of both object and scene representations, particularly for scenes, which had not been achieved previously. They perform a quantitative analysis on multi-colored dsprites using the ARI score to evaluate scene representation.
+
+### Strengths
+1. The authors present a unified approach to learning both object and scene representations through algebraic methods.
+
+2. The authors achieve high ARI scores of 0.99 and 0.98 on two different experiments, with the number of objects varying, indicating a strong performance in scene representation.
+
+### Weaknesses
+1. The paper lacks novelty, as most of the work is derived from Ohmura et al.'s work.
+
+2. The theory section is largely based on the inspired work, with limited original content.
+
+3. Equation 4, the loss function, is also a derivative of prior work.
+
+4. The experimental section is limited, with only two settings (two objects and different number of objects) and no comparison to other baselines.
+
+1. The authors do not provide an evaluation of object representation.
+
+2. The caption for Figure 2 is unclear and does not effectively convey the author's intentions.
+
+3. There is no other ablation provided on the independence of the approach.
+
+4. The authors do not explain the O transformation in the two experiments.
+
+5. It is difficult to understand the approach's benefits since there is no direct comparison with the SOTA methods. Further, the authors modified the multi-dsprites dataset (so no occlusions are present), making an even harder comparison with previous approaches. We suggest the authors train SOTA scene/object representation methods for the same setup.
+
+6. Another problem with evaluation is that the tested method was tested on only a simplified version of the multi-dsprites dataset. We encourage the authors to perform additional experiments on other datasets like CLEVR (a more challenging dataset than multi-sprites), and Tetrominoes.
+
+7. If the proposed method cannot handle cases where objects are partially occluded, the authors should clearly mention this in their paper as a limitation of their approach.
+
+8. The paper used segmentation accuracy as the quantitative metric. However, as the representations are derived from segmentation masks, segmentation accuracy cannot be used as a proper metric for evaluating representations. (Disentanglement metrics and other downstream tasks are needed)
+
+9. There are multiple attributes in the dataset while the paper only involves 2 sets of latent vectors and 2 transformations, which means the proposed method is limited in generalizing to real scenario.
+
+10. How to determine the correspondence between objects among the two scenes (how to determine the object pairs for transformation) is not clear.
+
+### Questions
+View the section above.
+
+### Soundness
+1 poor
+
+### Presentation
+2 fair
+
+### Contribution
+1 poor
+
+---
+
+## Human Reviewer 3
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+The paper proposed a unified theory to explain both scene representation learning and object representation learning based on algebraic independence and validated the theory in Multi-dSprites dataset.
+
+### Strengths
+The algebraic independence between object representation learning and scene representation learning that this paper focuses on is worth investigating.
+
+### Weaknesses
+The novelty of this work seems incremental, compared to Ohmura et al. (2023), it seems that the ALGEBRAIC INDEPENDENCE is straightforwardly implemented to scene and object representations. The novelty of this paper seems not that strong.
+
+The writing needs to be enhanced. There are some repetition in the ABSTRACT and INTRODUCTION sections, and the main contribution of the INTRODUCTION section of the paper is not very clear and concise. For example, the authors emphasized that they proposed a unified framework to explain scene and object representation learning, however, it seems difficult to follow that how  scene and object representation learning be explained.
+
+Regarding the UNIFIED THEORY section, the Object Representation Learning Description is a quotation from existing work[1] and lacks novelty. Overall, the idea of the proposed unified explanation framework is relatively simple and straightforward.
+[1] Yoshiyuki Ohmura, Wataru Shimaya, and Yasuo Kuniyoshi. An algebraic theory to discriminate qualia in the brain. arXiv preprint arXiv:2306.00239, 2023.
+
+### Questions
+In addition to the above weakness, I also have concerns about the experimental results.
+This paper consider the segmentation as the evaluation task, and uses a very traditional measurement adjusted Rand index (ARI) in evaluation. However, the widely used segmentation measurement of accuracy, mIoU are not considered. And there is no comparison to the related works in experiments, making it difficult to evaluate the advances of this paper.
+
+### Soundness
+3 good
+
+### Presentation
+2 fair
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 4
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+The paper attempts to present an approach to scene and object representation learning. Objects are decomposed into attributes and transformations between objects are learned using a neural network to satisfy algebraic independence between the attributes. A similar strategy is used to decompose scenes into objects and learn a transformation between scenes in the latent space. The authors claim the neural network used to learn these transformations accomplishes a form of representation learning.
+
+### Strengths
+Difficult to identify any contributions.
+
+### Weaknesses
+- The entire work seems like a trivial addendum to the cited work of Ohmura et. al. (2023) (which itself is a non peer reviewed paper on arXiv). Unfortunately, I was not able to identify significant extensions and significantly new experiments relative to Ohmura et. al. (2023).
+- The paper is poorly written:
+    - Vague explanations of technical concepts.
+    - Sentences are repeated verbatim in multiple places. E.G. abstract and introduction are word-for-word almost the same.
+    - Paragraph 3, line 2 seems to be copied from Ohmura et. al. (2023) page 6, line 1.
+    - The authors refer to other relevant work, but contributions relative to these references are poorly motivated and poorly explained. I was not able to find explanations how and why the other works relate to the proposed work.
+- I had difficulties in understanding the experiments, due to confusing setups and evaluation metrics.
+- Figs 1 and 2 do not explain the setup adequately; details missing.
+
+### Questions
+- Motivation and utility of this approach are poorly presented. How can we use the proposed neural networks for downstream tasks?
+- How would this approach handle real world images that contain complex scenes with overlapping and occluded objects?
+
+### Soundness
+1 poor
+
+### Presentation
+1 poor
+
+### Contribution
+1 poor

human_reviews/OwtkY3LyvC.md

@@ -0,0 +1,249 @@
+# Bridging The Gap Between Training and Testing for Certified Robustness
+
+- Decision: Reject
+- Scores: 3, 5, 1, 5, 3
+
+## Abstract
+Certified robustness provides a theoretical lower bound for adversarial robustness and arouses widespread interests and discussions from the research community. With theoretical support to improve the certified robustness on the training set, practitioners endeavor to train a more certified robust model during inference on the test set. However, the experimental neglect on the training set and the theoretical ignorance during inference on the test set induce a gap between training and testing for certified robustness. By establishing an equivalence between the convergence of training loss and the improvement of certified robustness, we recognize there is a trade-off between expressive power and generalization (assuming a well-conditioned optimization) for certified robustness, which is similar to the underfitting and overfitting discussed in machine learning. To investigate this trade-off, we design a new orthogonal convolution-Controllable Orthogonal Convolution Kernel (COCK) which provides a wider range of expressive power than existing orthogonal convolutions. Empirically, there is a power-driven shift from vanilla classification accuracy to certified robustness in the sense of the optimal trade-off between expressive power and generalization. The experimental results suggest that by carefully improving the expressive power from the optimal trade-off for vanilla classification performance, the model will be more certified robust.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+3
+
+### Summary
+This paper presents an analysis on certified robustness on Convolutional Netural Networks (CNNs). The authors first show an equivalence between the convergence of training loss and improvement of certified robustness. After that, a new orthogonal convolution-Controllable Orthogonal Convolution Kernel (COCK) is proposed to mitigate the trade-off between expressive power and generalization. Through experiments on CIFAR10 dataset, the authors show that COCK achieves improvement on vallina accuracy and certified accuracy compared to existing baselines.
+
+### Strengths
+The proposed method performs better than the baseline on most cases.
+
+### Weaknesses
+I believe the current paper is not in a publishable quality for the following reasonings:
+1. Preliminary is not clear. In Section 3, the definition of cerfified robustness is missing. It's also not well-explained how it's connected to orthogonal convolution and Lipschitz-contrainted models.
+2. Definition is not clear, notation is not consistent. For example, in the beginning of Section 3, $x^{(i)}$ denotes the $i$-th example in the training set, while in Equation (10), $x^{(ij)}$ denotes the $j$-th example that belongs to class $i$. The definitions of Convolition (3.3.1), margin loss (4.2.1) and certified robust accuracy (4.2.3 and 4.2.4) are also not clear.
+3. Theorem 4.1.1 is not clear. I don't think it trivial that "minimizing training loss is equivalent to maximizing the output margin $m_k^{(ij)}$". It's also not clear how equivalent it is between the derease of training loss and the improvement of certified robustness. One should give a mathematical description to support the argument. 
+4. The proposed algorithm COCK is not well explained. How it it's related to Proposition 5.1.1 (and Fourier Transform)？What does COCK actually do? There is not adequate explanation on section 5.
+5. How COCK is related to r-RELU? Also, Figure 1 is not well captioned.
+6. The experiments are limited to CIFAR10 dataset and LipConvNet. Also, the citation of baselines (SOC, BCOP and Cayley) are missing.
+
+### Questions
+Please refer to the "Weakness" above.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+1
+
+---
+
+## Human Reviewer 2
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+2
+
+### Summary
+This paper aims to understand the differences in certified robustness observed during training and testing phases, especially focusing on how the expressive power of models impacts certified robustness on the test set.
+
+The authors identify a trade-off between expressive power and generalization for certified robustness, drawing an analogy to the overfitting and underfitting dynamics in traditional machine learning. To explore this, they introduce a new model component, the Controllable Orthogonal Convolution Kernel (COCK), which allows for greater flexibility and a broader range of expressive power than existing orthogonal convolution methods.
+
+### Strengths
+1. This paper provides a fresh perspective by drawing a parallel between the traditional machine learning concepts of underfitting and overfitting and the balance between expressive power and generalization in certified robustness. This analogy sheds light on an often-overlooked aspect in the field, emphasizing that certified robustness exhibits similar trade-offs as seen in conventional learning. This insight could potentially open new avenues for the research community to deepen their understanding of certified robustness.
+
+2. The authors introduce the Controllable Orthogonal Convolution Kernel (COCK), a module specifically designed to adjust the expressive power of neural networks. By enabling more nuanced control over the model’s expressive capacity, COCK serves to bridge the gap between training and test robustness.
+
+### Weaknesses
+1. The central finding of the paper—that there exists a trade-off between expressive power and generalization for certified robustness— is not surprising and may be considered common knowledge in this field. While the paper empirically validates this trade-off, it lacks both theoretical understanding and empirical analysis of this trade-off.
+
+2. Although the paper introduces new definitions and theorems, these contributions primarily reinterpret existing ideas rather than advancing theoretical understanding. For instance, Proposition 5.1.1 on COCK appears to be more of an application or extension of existing techniques rather than a novel theoretical innovation. Thus, the paper’s contributions seem limited.
+
+### Questions
+See the Weaknesses. 
+
+Overall, I think this paper raises an interesting phenomenon, but the theoretical research or experimental analysis of this phenomenon is not deep enough and needs further improvement.
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 3
+
+### Rating
+1
+
+### Rating Number
+1
+
+### Confidence
+4
+
+### Summary
+This paper designs a new certified robustness method named COCK that is claimed to enjoy a better robustness-utility trade-off (but I don't understand what is the intuition behind the design of COCK). Experiments are conducted on CIFAR-10, CIFAR-100, and Tiny-ImageNet to evaluate the proposed COCK method (which shows that COCK is significantly worse than most recent certified robustness methods).
+
+### Strengths
+None.
+
+### Weaknesses
+1. The certified robustness performance of the proposed COCK method is significantly weaker than recent certified robustness methods such as [r1] and [r2]. Specifically, for example, according to Table 1, when the $l\_2$-perturbation radius is $\frac{108}{255}$, the best certified robust accuracy COCK achieved on CIFAR-10 dataset is $60.79\\%$. However, with a higher perturbation radius $\rho = 0.5 > \frac{108}{255}$, on the CIFAR-10 dataset, [r1] and [r2] achieved certified robust accuracies of $65.5\\%$ and $70.7\\%$ respectively, which both are significantly higher than the result of COCK. Similar things also happen for the other dataset (CIFAR-100) and other perturbation radii. Furthermore, [r1] and [r2] can work on the ImageNet dataset, while COCK is only analyzed on the Tiny-ImageNet dataset. So if the proposed COCK method is so weak, why the community needs to waste time on it?
+
+2. Algorithm 1 (i.e., COCK) seems to be very time-consuming, which may significantly limit the practicality of the proposed COCK method. There are two main calculation challenges: (1) Line#681 needs to conduct $3t$ times of matrix multiplication (i.e., calculate the term $B^3\_{\tau-1} S$ for $t$ times), and (2) Line#688 needs to perform inverse-FFT for $i*j$ times. So I suggest the authors analyze the theoretical and empirical time complexity of COCK. Also please note that [r1] and [r2] can even work without additional model training.
+
+3. The authors claim that one of their contributions is that they discover "a trade-off between expressive power and generalization for certified robustness" (For example, see Line#078-079). Unfortunately, they are not the first to find the trade-off and such an observation is now very trivial. Specifically, the trade-off between model robustness and utility has been widely accepted by the ML S&P community and has also been well studied in many existing works such as [r3], [r4], [r5], and [r6].
+
+4. I still do not understand what is the motivation for designing COCK? How can it make COCK enjoys certified robustness better than existing certified robustness approaches (although it actually cannot)? Please comment.
+
+5. Experiments could be significantly improved. Firstly, I suggest including experiments on the (vanilla) ImageNet dataset. Secondly, I think the authors should cite more recent certified robustness baselines and compare them with your proposed COCK. It is worth noting that the most recent paper that the authors cited in this work is from 2022, which is still a too-old paper (in the field of certified robustness).
+
+**References:**
+
+[r1] Carlini N. et al. "(Certified!!) Adversarial Robustness for Free!" arXiv 2206.10550 / ICLR 2023.
+
+[r2] Chen H. et al. "Diffusion Models are Certifiably Robust Classifiers." arXiv 2402.02316 / NeurIPS 2024.
+
+[r3] Raghunathan A. et al. "Understanding and Mitigating the Tradeoff between Robustness and Accuracy." ICML 2020.
+
+[r4] Donhauser K et al. "Interpolation can hurt robust generalization even when there is no noise." NeurIPS 2021.
+
+[r5] Rice L. et al. "Overfitting in adversarially robust deep learning." ICML 2020.
+
+[r6] Fu S. et al. "Theoretical Analysis of Robust Overfitting for Wide DNNs: An NTK Approach." ICLR 2024.
+
+### Questions
+See **Weaknesses** for details.
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+---
+
+## Human Reviewer 4
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+3
+
+### Summary
+This manuscript introduces the controllable orthogonal convolutional kernel (COCK), which uses a more flexible parameterization of convolution operation while controlling its Lipschitz constants. In this context, COCK demonstrates better flexibility while ensuring certified robustness.
+
+### Strengths
+1. The gap of certified robustness between the training and the test set is a key issue in provably robustness while underexplored. The proposed method is theoretically guaranteed.
+
+2. The experiment is comprehensive.
+
+3. The manuscript is generally well-written and easy to follow.
+
+### Weaknesses
+1. My major concern is the logic of the motivation and the proposed method. As Section 5 demonstrates, the proposed COCK is more flexible than vanilla orthogonal convolution, why doesn't it suffer more from overfitting? The last paragraph in Section 4.1 points out that there is a trade-off between expressive power (related to parameterization flexibility) and generalization.
+
+2. Theorem 5.2.1 may need more explanation. I agree COCK demonstrates no-inferior performance in the experiment, but I am suspicious about the proof of this theorem. In the proof, especially Equation (35), the authors seem to assume that the margins of COCK and the baseline are the same, while it may not be the case. In addition, when using COCK for parameterization, there may exist no-inferior points compared with the baseline, but we cannot guarantee the optimization in training and lead the parameter to converge to such cases. In this regard, the conclusion of Theorem 5.2.1 looks too strong, the authors may need more assumptions or explanations to make this part more rigorous.
+
+3. Although the experiment is comprehensive, the $\rho$ that we achieve good certified robustness is small, limiting the practical use of the method. Actually, in the empirical study of adversarial training, the $\epsilon$ value in the $l_2$ cases is usually set as $0.5$ or $2$, much larger than the values used in the experiment (e.g. Table 1) of this manuscript.
+
+### Questions
+The major questions are pointed out in the "weakness" part, the authors need to answer those questions to address my concerns.
+
+Two minor issues:
+
+1. In Definition 4.2.2, the difference between $\epsilon$ and $\epsilon_2$ seems unclear.
+
+2. In Figure 1, the authors should distinguish the difference between the original input and the post-transformation features. In the current figure, it looks like there are $8$ points in total.
+
+From my point of view, the current version of the manuscript is not ready for publication. I welcome the authors to address my concerns in the rebuttal.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+---
+
+## Human Reviewer 5
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+The paper studies the gap between training and test certified accuracy and identifies a questionable trade-off between expressive power and generalization. Following this, the paper proposes a new orthogonal convolution named Controllable Orthogonal Convolution Kernel (COCK) which improves the certified robustness.
+
+### Strengths
+1. The paper studies an important problem of certifying robustness in neural networks.
+2. The new convolution COCK learns parameters in the frequency domain which is novel and seems to improve the certified robustness.
+
+### Weaknesses
+The argument for the trade-off between training and test certified robustness raises some fundamental issues that are not discussed in the work (see Questions).
+
+### Questions
+### The fundamental questions I have are the following:
+1. The trade-off between expressive power and generalization for certified robustness is questionable. It is primarily established based on the argument that "expressive power brings high generalization error to classification performance during inference (overfitting) and is certain to degrade the certified robustness on the test set" (lines 245-247). But, this doesn't hold for commonly used overparameterized neural networks, as they exhibit double descent phenomenon, that is, increasing the expressive power (near 0 training error) decreases the generalization error (test error) [P1]. 
+2. Why the theoretical results established on the training set cannot extend to the test set from the same data distribution? The argument made by the authors for expecting a gap between the training and test certified robustness is "experimental neglect on the training set and the theoretical ignorance during inference on the test set" -- but, statistically when training and test sets are from the same distribution, the theory established on large training data would translate to the test set as well. 
+
+### Questions on COCK:
+1. In Algorithm 1 (line 670), is the inverse Fourier transform in line 688 new compared to the existing orthogonal convolutions? 
+2. Line 295: it is not clear why $h=w=n$ and what is $n$? 
+3. Line 324: how does COCK provide data scaling? Please clarify it.
+4. Line 330: in the toy example case, why is the comparison made against a diagonal matrix? Would COCK result in a diagonal matrix in any of the settings?
+5. The toy example visualization is unclear. What exactly is visualized? Are the points in data space, or are they layer wise feature space? The feature space dimension is also not mentioned. 
+
+### Questions on experiments:
+1. How COCK is different from SOC, BCOP, and Cayley, the orthogonal convolutions used in experiments? Please add citations for the other algorithms and briefly compare and contrast them. 
+2. The closest algorithm to COCK is layer-wise orthogonal training (LOT), as mentioned in line 116. However, a direct comparison is *not* made experimentally. Are the orthogonal convolutions used (SOC, BCOP, Cayley) do layer-wise orthogonalization? 
+3. what is max-min activation mentioned in line 383?
+4. Line 393: $\rho$ is used for l2 norm perturbation. I think it is the same as $\epsilon$ used in sections before. Is it? If so, please follow the same notation for clarity.
+
+Overall, I think the new convolution is novel and interesting that it improves the certified robustness compared to the other orthogonal convolutions. But, the premise of the trade-off analysis is not convincing. 
+
+[P1] Belkin et al. "Reconciling modern machine-learning practice and the classical bias–variance trade-off." PNAS 2019
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2

human_reviews/PEuDO2EiDr.md

@@ -0,0 +1,178 @@
+# RTFS-Net: Recurrent Time-Frequency Modelling for Efficient Audio-Visual Speech Separation
+
+- Decision: Accept
+- Scores: 8, 8, 6, 8
+
+## Abstract
+Audio-visual speech separation methods aim to integrate different modalities to generate high-quality separated speech, thereby enhancing the performance of downstream tasks such as speech recognition. Most existing state-of-the-art (SOTA) models operate in the time domain. However, their overly simplistic approach to modeling acoustic features often necessitates larger and more computationally intensive models in order to achieve SOTA performance. In this paper, we present a novel time-frequency domain audio-visual speech separation method: Recurrent Time-Frequency Separation Network (RTFS-Net), which applies its algorithms on the complex time-frequency bins yielded by the Short-Time Fourier Transform. We model and capture the time and frequency dimensions of the audio independently using a multi-layered RNN along each dimension. Furthermore, we introduce a unique attention-based fusion technique for the efficient integration of audio and visual information, and a new mask separation approach that takes advantage of the intrinsic spectral nature of the acoustic features for a clearer separation. RTFS-Net outperforms the prior SOTA method in both \textbf{inference speed} and \textbf{separation quality} while reducing the number of parameters by \textbf{90\%} and MACs by \textbf{83\%}. This is the first time-frequency domain audio-visual speech separation method to outperform all contemporary time-domain counterparts.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+In this paper, the authors present a novel time-frequency domain audio-visual speech separation method (Recurrent Time-Frequency Separation Network), a unique attention-based fusion technique for the efficient integration of audio and visual information, and a new mask separation approach. Results show that the proposed approach outperforms the previous SOTA method using only 10% of the parameters and 18% of the MACs. The authors claim that this is the first time-frequency domain audio-visual speech separation method to outperform all contemporary time-domain SOTA ones.
+
+### Strengths
+One main strength of this paper is the proposed Recurrent Time-Frequency Separation Network that processes the data in the frequency dimension, the time dimension, and the joint time-frequency dimension.
+
+### Weaknesses
+It is mentioned that the RTFS blocks share parameters (including the AP Block), leading to reduced model size and increased performance. Therefore, more description/explanation for this would be helpful.
+
+### Questions
+Are there any overlapping speech in the train/test data?
+Do the authors perform any downstream task like speech recognition on the reconstructed speech?
+One possible downstream task for speech separation is speech-to-speech dubbing. In this case, both the speech and background/nonspeech sound are needed. Have the authors looked into reconstructing background/nonspeech sound?
+
+### Soundness
+4 excellent
+
+### Presentation
+4 excellent
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 2
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper proposes RTFS-Net, a new time-frequency (TF) domain audio-visual speech separation method. It introduces three main innovations:
+
+* RTFS Blocks independently model time and frequency dimensions of audio
+* Cross-dimensional Attention Fusion Block efficiently fuses audio and visual data
+* Spectral Source Separation Block preserves phase/amplitude information
+
+Experiments show RTFS-Net matches or beats prior time domain methods on LRS2, LRS3, and VoxCeleb2 datasets, while using 10x fewer parameters and 3-6x fewer computations.
+
+RTFS-Net is the first TF model to surpass most contemporary time domain methods for audio-visual speech separation. It demonstrates TF domain methods can achieve good performance at lower computational cost through novel modeling of time-frequency spectrograms.
+
+### Strengths
+* Achieves near state-of-the-art performance for audio-visual speech separation while being very parameter and computationally efficient
+* Outperforms all compared time domain methods, proving time-frequency domain modeling can achieve better performance if done effectively
+* Innovative modeling of time and frequency dimensions independently in RTFS Blocks
+* Attention-based fusion mechanism in CAF Block is very lightweight but fuses audio and visual data very effectively
+* Spectral Source Separation Block properly handles phase/amplitude to avoid losing audio information
+* Model code and weights will be released for full reproducibility
+
+### Weaknesses
+* Testing is limited to only 2 speaker mixtures. Performance with more speakers is uncertain.
+* Missing PESQ evaluation in results which most other target speech extraction papers provide
+* Doesn't include latest SOTA model comparison: Dual-Path Cross-Modal Attention for Better Audio-Visual Speech Extraction, ICASSP 2023.
+https://arxiv.org/pdf/2207.04213.pdf. This gives superior performance for SI-SNRi and provides PESQ results as well. It does not provide MACs analysis.
+
+### Questions
+None
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 3
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The authors build upon previous research in audio-only and audio-visual speech recognition by focusing on improving efficiency and fidelity of separated speech. They draw a lot of inspiration from the CTCNet paper and extend it to the TF domain to improve the efficiency. The solution has been evaluated on standard benchmark datasets and compared to previous state-of-the-art methods.
+
+### Strengths
+1. Audio samples of separation are available and source to be made available when the paper is published.
+2. The writing is easy to follow.
+3. Clear modeling details are provided.
+
+### Weaknesses
+1. The baseline methods listed in table 1 should include their references.
+2. RTFS-Net-12 is only about 10% more efficient that CTCNet. How much difference does that make in practical applications?
+3. Some of the comparison examples are not distinguishable to this reviewer. This makes me wonder how to interpret the relative SNR gains.
+
+### Questions
+1. Since the best performance is achieved with R=12, why not explore a higher R?
+2. Have the authors considered conducting studies with human listeners? If the target application is ASR, would it be helpful to measure WER in a recognition task?
+
+### Soundness
+3 good
+
+### Presentation
+2 fair
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 4
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper presents a novel Recurrent Time-Frequency Separation Network architecture that performs audio-visual source separation tasks effectively and efficiently. The model is characterized by three parts. First, each modality goes through its own processing module, and then the cross-dimensional attention fusion (CAF) consolidates information from both modalities. The spectral source separation block performs masking-based separation. The separation results show promising improvement given the compact size and computational efficiency the new model architecture introduces.
+
+### Strengths
+- The paper presents solid improvement compared to the existing baseline systems. Considering the amount of model compression the proposed model introduced, the improvement is significant.
+
+- All the procedures and modules are well-defined with enough details.
+
+- The choice of the model architectures makes sense, including the dual-path structure, attention-based consolidation, and complex masks.
+
+- Ablation studies are thorough.
+
+### Weaknesses
+While the paper is packed with useful information, there are still some parts that need elaboration.
+
+- As the authors mention, the dual-path RNN idea is not new to this problem. I understand that the authors chose SRU for their complexity-related considerations, but I also wonder if the audio processing module could benefit from its own self-attention mechanism, such as in the SepFormer model. 
+
+- The spectral source separation module might be the weakest contribution, because complex masks have been extensively studied in the audio-only source separation literature. 
+
+- I wish the paper provides more details on the TDANet block for video processing, which is relegated to the reference in the current version.
+
+### Questions
+- The authors chose to "add" f_1 and f_2 (eq 11) after the CAF processing. I think it's a little abrupt in the sense that there might be other choices that preserve the unique information that each vector learns, such as concatenation. Have the authors considered other ways to combine the two vectors?
+
+### Soundness
+4 excellent
+
+### Presentation
+3 good
+
+### Contribution
+4 excellent

human_reviews/RFjhxXrTlX.md

@@ -0,0 +1,154 @@
+# Learning Unorthogonalized Matrices for Rotation Estimation
+
+- Decision: Reject
+- Scores: 3, 6, 5
+
+## Abstract
+Estimating 3D rotations is a common procedure for 3D computer vision. The accuracy depends heavily on the rotation representation. 
+One form of representation -- rotation matrices -- is popular due to its continuity, especially for pose estimation tasks. The learning process usually incorporates orthogonalization to ensure orthonormal matrices. Our work reveals, through gradient analysis, that common orthogonalization procedures based on the Gram-Schmidt process and singular value decomposition will slow down training efficiency. 
+To this end, we advocate removing orthogonalization from the learning process and learning unorthogonalized `\textbf{P}seudo' \textbf{Ro}tation \textbf{M}atrices (PRoM).
+An optimization analysis shows that PRoM converges faster and to a better solution. By replacing the orthogonalization incorporated representation with our proposed PRoM in various rotation-related tasks, we achieve state-of-the-art results on large-scale benchmarks for human pose estimation.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.
+
+### Summary
+This paper claims that a popular rotation estimation technique (6D representation) for deep learning has a critical flaw, i.e., ambiguous gradient. The paper tries to demonstrate this with some derivations, and the authors propose to use the pseudo rotation matrix technique (using a plain 3x3 matrix directly, with an additional loss that makes it similar to the ground truth rotation) to overcome this issue. This pseudo rotation matrix is orthogonalized in the test phase. Experiments show that the proposed method has less variation in the gradient, and it shows somewhat better performance in several downstream tasks, including 3D body/hand pose and shape estimation.
+
+### Strengths
+The topic of the paper (rotation estimation in deep learning) has some importance in the field.
+
+### Weaknesses
+- To tell the truth, I was one of the reviewers of this paper in a previous venue. I see that the authors have removed some wrong derivations, but there are still many vague and wrong claims in the paper.
+
+- Most claims in the paper are either vague or not surprising (not showing what the authors intended to show). Most importantly, the proposed method (using a plain 3x3 matrix without any orthogonalization and instead guiding it to a rotation matrix using an additional loss, i.e., soft constraint) is a somewhat basic technique that has been frequently used in the literature. Accordingly, I'm afraid the paper has little merit to the research community. The detailed errors (or concerns I have) are listed below:
+
+- Section 3.4.1 (and Appendix A.3) basically says that (5) has two more terms other than the first term, so the gradient can be scrambled. The claim itself is vague, but before that, what the authors need to show is not that there are two more terms, but that these two terms do not die out (and actually steer the overall gradient towards funny directions) in the majority of the input space. In (12)-(16), we can actually see that there are many duplicating expressions, as well as many terms related to cross-product, inside those three terms. This suggests that the two terms (related to $r_2$ and $r_3$) might actually be canceled, either by themselves or by combining the two. I actually showed this in my previous review for a specific case provided in the previous manuscript. This might hold for larger regions in the input space if we do the manipulation. The authors must explicitly derive against this to justify their claim, not suggesting them as in the current paper.
+
+- Theorems in Section 4.2 are not surprising, and most importantly, they do not prove anything. What they are showing are (specific) upper bounds of losses for orthogonalized and unorthogonalized cases. This does not guarantee anything. They do not guarantee that the actual loss will be lower for the unorthogonalized case, and moreover, there is no proof that they are the best upper bounds we can have (so it is not correct to say that the unorthogonalized case has a better chance of a lower loss).
+
+- The meaning of Fig. 3 is vague. It simply shows that the correlation between the gradient and $t_{11}$ is not apparent in Euclidean space. This has nothing to do with the stability or ambiguity of the gradient. More importantly, it is obvious that the relation between the gradient and the parameter is more complex for orthogonalization than for plain identity mapping.
+
+- What if we do not have ground truth rotations? Without $L_\theta$, it is possible that PRoM deviates too far from the rotation space, and this can be detrimental to having a correct solution. What should we do then?
+
+### Questions
+Please see the above weaknesses.
+
+### Soundness
+1 poor
+
+### Presentation
+2 fair
+
+### Contribution
+1 poor
+
+---
+
+## Human Reviewer 2
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+The work studied the rotation representation for rotation estimation and other downstream tasks, such as human body and hand shape/pose estimation. An issue with orthogonalization in learning was pointed out: this could lead to explosive gradient and hence harm the training stage stability. To overcome this, this work proposes to get rid of the step of orthogonalization during learning, and only including this step during inference step. Theoretical proofs are derived to support the claim of the proofs and evaluations on multiple tasks verify the effectiveness of the proposed rotation representation.
+
+### Strengths
+1. The work focus on the choice of rotation representation, the very basic but vital ingredient in rotation estimation task. This is in contrast to most work in 3D pose estimation, and has the potential of larger impact in the field. 
+2. The method is simple yet effective. The motivation for the method is put clear. Gradient update is affected by the extra step of orthogonalization. By simply removing it, the baseline method could be improved. 
+3. Evaluations tasks are very extensive. These include human body, hand, and pure rotation estimation tasks. Three cases are considered with respect to the availability of intermediate rotation ground truth / end-task supervision.
+
+### Weaknesses
+1. The evaluations consider two baselines (Zhou et al., 2019; Levinson et al., 2020). Both methods use 6D representations with differences in normalization. A comparison to more recent methods [1,2] should be more convincing. 
+
+2. The proposed method chose to base on CLIFF (Li et al., 2022). Since the proposed method is not a novel framework, it is generally applicable to any existing methods for pose estimation. Hence, I assume that PROM could be used together with baselines to also improve learning. 
+
+3. In the area of camera pose regression, it has been shown that predicting pseudo rotation matrix during training works in [3] and quantanion without orthogonalization is also possible [4]. As a result, the novelty of this work might be reduced to introducing the simple idea to other tasks. 
+
+
+[1] Projective Manifold Gradient Layer for Deep Rotation Regression, Jiayi Chen etc. 
+[2] Deep Projective Rotation Estimation through Relative Supervision, Brian Okorn etc. 
+[3] Direct-PoseNet: Absolute Pose Regression with Photometric Consistency, Shuai Chen, etc. 
+[4] Posenet: A convolutional network for real-time 6-dof camera relocalization. Alex Kendall, etc.
+
+### Questions
+1. There may be a typo in Sec 3.4.1. In Eq.5, the gradient is respect to $t'_1$, but the text refers to $t'_2$ and $t'_3$ to show the column-gradient will be in different directions. Please confirm this. 
+2. This method seems to be very simple, simply removing the $r$ and $g$. why is removing $r$ possible? If I understand correct, $r$ is the mapping from input to rotation representation, and should not be identity mapping. 
+3. My biggest concern is: if orthogonalization can be safely removed as suggested by this work, why the previous work adopted GS or SVD to make the rotation representation valid?
+
+### Soundness
+3 good
+
+### Presentation
+2 fair
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 3
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper presents PRoM, a method to learn pseudo-rotation matrices. The motivation is that the paper claims that learning rotation matrices in a deep learning framework suffers from imposing orthonormality constraints by means of SVD or Gram-Schmidt. The proposed method removes these orthogonalization methods and apply the estimated rotation matrix. The paper presents a mathematical analysis discussing the reasons about why SVD or Gram-Schmidt suffer when applied in a deep learning framework. The paper presents experiments on various applications (e.g., human pose estimation and point cloud pose estimation) showing that PRoM outperforms the included baselines.
+
+### Strengths
+S1. I think the paper does a good job analyzing the problems that orthogonalization methods introduce when used in the learning frameworks.
+
+S2. I find the experiments interesting because they show the improvements PRoM can bring to human pose estimation, cloud pose estimation, among other applications.
+
+S3. The paper is well written and is easy to follow  and understand. Clarity thus is high and should be easy to replicate.
+
+### Weaknesses
+While I think the paper touches a very interesting topic (i.e., rotation estimation in deep learning frameworks) and shows interesting results, I have several concerns that make me a bit skeptical about the approach:
+
+1. Lack of discussion about other possible ways to constrain orthonormality in learning problems. While the problems of orthogonalization in the learning problem may bring issues as described in the paper, there are other possible solutions that the paper does not discuss. For example: 
+
+i) In 3D computer vision, some camera-pose estimation solvers constrain the rotation with $I - M^{\intercal}M$ and $\text{det}(M)=1$ pose problems and showing good estimation results (see Reference [A]). Why couldn't this trick replace the SVD or Gram-Schmidt orthogonalization procedures? Although, the paper does not discuss a learning approach, the trick is applicable to any learning problem. Unfortunately, the paper lacks discussion about these tricks and if they also present problems when learning.
+
+ii) Learning rotation with neural networks does not need to follow the pipeline shown in Figure 1. One can easily use a quaternion parametrization, remove $\mathcal{L}_{\theta}$ and thus remove the orthogonalization, and simply add a regularizer term that enforces the norm of the quaternion be one.
+
+2. The continuity issues discussed in the submission and that are used to motivate the problem are questionable. One of the main claims of the paper in Learning for Rotations in Section 2 states that the parametrization of 3D rotations with four or fewer dimensions is discontinuous and non-ideal for learning (according to Zhou et al. 2019). However, Zhou et al. 2019 defines the continuity concept from an intuitive example and lacks solid theorems showing that these definitions and observations strictly apply to learning rotations. Subsequent works using quaternions show that they work well with neural networks (see Reference [B] below). Reference B conflicts with the observations of Zhou et al. which is the foundation of this work and explore pseudo-rotation matrices. Thus, this puts into question the main assumptions stated in Section 2.
+
+3. The paper would've been more solid if they included the discussed methods in 1 (see above).
+
+References.
+[A] Ventura, Jonathan, et al. "A minimal solution to the generalized pose-and-scale problem." CVPR. 2014.
+[B] Zhang et al. Quaternion Product Units for Deep Learning on 3D Rotation Groups. CVPR 2020
+
+### Questions
+Given the rationale described in the Weaknesses above, I don't find the proposed method of using pseudo-rotation matrices that convincing. It marries to the pipeline of learning rotations to that one described in Figure 1 of the submission. However, I think there are other ways to learn rotations that are not discussed in the submission.
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+2 fair

human_reviews/SYPx4NukeB.md

@@ -0,0 +1,159 @@
+# SSL Framework for Causal Inconsistency between Structures and Representations
+
+- Decision: Reject
+- Scores: 3, 5, 6
+
+## Abstract
+The cross-pollination of deep learning and causal discovery 
+  has catalyzed a burgeoning field of research, 
+  seeking to elucidate causal relationships within 
+  non-statistical data forms like images, videos, and text. 
+  Such data, often being named `indefinite data',
+  exhibit unique challenges—inconsistency between causal structure 
+  and representation, which are not common in conventional data forms.
+  To tackle this issue, we theoretically develop intervention 
+  strategies suitable for indefinite data and derive 
+  causal consistency condition (CCC). Moreover, 
+  we design a self-supervised learning (SSL) framework 
+  that considers interventions as  `views' and CCC 
+  as a `philosophy' with two implement examples on 
+  Supervised Specialized Models (SSMs) and Large Language Models 
+  (LLMs), respectively. To evaluate pure inconsistency manifestations, 
+  we have prepared the first high-quality causal dialogue dataset-
+  \textit{Causalogue}. Evaluations are also performed 
+  on three other downstream tasks. Extensive experimentation 
+  has substantiated the efficacy of our methodology, 
+  illuminating how CCC could potentially play an influential role 
+  in various fields. Our code is available in~\href{https://anonymous.4open.science/r/ICLR_Anonymous_submission_575_new_dataset_Causalogue_and_codes/Causalogue_Dataset/Fork_I_demo.py}{url of 
+  anonymous code and data}.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+The paper proposes a causally motivated self supervised learning framework. In particular, the paper focuses on indefinite data, which refers to data that requires deep networks to represent them like text, videos, images etc. The authors show that current methods do not learn consistent structure and representations from a causal perspective, while their proposed method based on causal consistency outperforms existing SOTA methods on different benchmarks
+
+### Strengths
+The idea of considering consistency between representation and structure as a way of learning causal relations in an unsupervised manner is interesting.
+
+### Weaknesses
+I feel the paper is poorly written. Without the Appendix, the technical contributions of the paper is very difficult to understand. Ideally, the paper should be complete with the Appendix dedicated to extra information that complements the contents of the paper. The experimental setup is not clearly explained which makes appreciating them difficult. More details in questions.
+
+### Questions
+1. The paper studies the inconsistency problem between causal structure and representation. However, what is structure and representation is never formally defined in the paper. I understand from Definition 1, structure refers to the graph and representation is simply the output of a deep network operating on the data, like word embeddings of text. However, the authors consider two causal models one with the representation and one with the structure, which is confusing. A concrete example with the same data, but the two different U and V causal models for the structure and the representation would go a long way in making the presentation more clear.
+
+2. Hypothesis 2 is not clear. What does the notation E(\hat{x}_{s,m,n}) = x_{s,m,n} mean? The E operator has not been defined, is it the expectation? If so, what is the expectation over?
+
+3. Section 2.2 is incomprehensible. What is causal consistency has not been defined so far, What is been plotted in Figure 1. The caption says it is the MSE between similarity matrices of representation and structure. What are these similarity matrices? Appendix A.3 does not give these details. What is the reconstruction loss the authors are referring to here? For example, lines 130-133, the authors say in the M > 1 or D > 1 case, the optimization of causal strength f changes to a weighted linear combination of f_m for the different M structures. However, this optimization problem has never been defined. Numerous such issues plague the readability of the paper.
+
+4. Eq 3 is unclear. What is L_k, what is being optimized? I understand at a high level we have 2 causal models, we intervene on both of them, and then have a way of checking the consistency of the two models. But beyond this high-level details, exact specifics of how the authors carry out the interventions and check for consistency is not present in the paper. 
+
+5. The authors propose a Causalogue dataset for causal discovery. However, it is not clear how this dataset was construcuted. The appendix gives some details. However, the authors handcraft 10 causal structures. It is not clear how GPT-4 was prompted to respect this causal structure while generating the utterances. 
+
+I think the entire paper needs significant restructuring to make the presentation clear and allow readers to understand the main contributions and needs more details to reproduce the results.
+
+### Soundness
+2 fair
+
+### Presentation
+1 poor
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 2
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper performs research on the intersection of deep learning and causal discovery. By so-called causal consistency, it proposes a self-supervised learning framework which provides potential support for the LLM. with some test results on a new dataset.
+
+### Strengths
+1. The idea is novel, and the viewpoint of consistency is important.
+2. The proposed new dataset may serve the community.
+
+### Weaknesses
+1. Some definitions need to be more accurate.
+2. Experimental section can be improved.
+
+### Questions
+1. The definition of "non-statistical data" forms like images, videos, and text is vague. These data still contains statistical information, and what do you mean by "non-statistical"?
+2. Definition 1 (Causal Data). I still think the definition of "Indefinite Data" "Semi-definite" based on D and M only is weak. How to quantify causal consistency is still an open question. This also applies to section 2.3.
+3. Def 3. The symbols looks slightly wired. 
+4. Section 6. I am still puzzled that how you evaluate the "causal graph" and "structure" both. It seems that these two things basically align with the same aspect of the algorithm. Why not evaluate "variance of learning results" or ”scalability of the algorithm“？
+
+### Soundness
+3 good
+
+### Presentation
+4 excellent
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 3
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+In this paper, the author proposes a SSL framework to make the causal structure be the same as causal representation by regulating their strength set. The author further introduces a dialogue dataset which could serve as a potential benchmark.
+
+### Strengths
+1. The author showcases implementations under different frameworks. 
+
+2. The dataset could be useful. 
+
+3. The experimental results are promising.
+
+### Weaknesses
+1. The presentation and the readability of the paper can be improved. 
+
+The author points out the inconsistency between causal structure and causal representation.  However, there is no formal definition of causal consistency in the paper.  In theory 1, the author gives the causal consistency condition as if it is equivalent under any intervention, these two causal models are consistent. But still, it is not a definition and there is a lack of intuition and motivation of why inconsistency is a problem to be solved. 
+
+
+2. The hypothesis 2 is confusing. 
+
+In images or text, people project the original data into latent space, where each dimension does not have to be entangled. Even in the feature space, they do not have to be entangled. The pixel one may not relate to pixel two.
+
+### Questions
+1. What is the causal inconsistency between causal representation and causal structure and why do we want to optimize it. 
+
+
+2.  How do you adapt your method to other datasets where the intervention cannot be done, for example, celebA.
+
+### Soundness
+3 good
+
+### Presentation
+2 fair
+
+### Contribution
+2 fair

human_reviews/TjfXcDgvzk.md

@@ -0,0 +1,179 @@
+# NOLA: Compressing LoRA using Linear Combination of Random Basis
+
+- Decision: Accept
+- Scores: 6, 6, 6, 6
+
+## Abstract
+Fine-tuning Large Language Models (LLMs) and storing them for each downstream task or domain is impractical because of the massive model size (e.g., 350GB in GPT-3).
+Current literature, such as LoRA, showcases the potential of low-rank modifications to the original weights of an LLM, enabling efficient adaptation and storage for task-specific models. These methods can reduce the number of parameters needed to fine-tune an LLM by several orders of magnitude. Yet, these methods face two primary limitations: (1) the parameter count is lower-bounded by the rank one decomposition, and (2) the extent of reduction is heavily influenced by both the model architecture and the chosen rank. We introduce \nola{}, which overcomes the rank one lower bound present in LoRA. It achieves this by re-parameterizing the low-rank matrices in LoRA using linear combinations of randomly generated matrices (basis) and optimizing the linear mixture coefficients only. This approach allows us to decouple the number of trainable parameters from both the choice of rank and the network architecture. We present adaptation results using GPT-2, LLaMA-2, and ViT in natural language and computer vision tasks. \nola{} performs as well as LoRA models with much fewer number of parameters compared to LoRA with rank one, the best compression LoRA can archive. Particularly, on LLaMA-2 70B, our method is almost 20 times more compact than the most compressed LoRA without degradation in accuracy
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+Low Rank Adaptation (LoRA) presents a series of drawbacks, particularly its constrained parameter reduction due to rank-1 matrices, which cannot be further diminished. Additionally, LoRA's parameter count is heavily reliant on the model's architecture. In response to these limitations, this paper suggests an innovative solution by advocating the use of a linear combination of random projections to replace LoRA's update matrix, effectively addressing the issues mentioned earlier. This approach is inspired by the previous paper known as PRANC. Personally, I found the paper to be a valuable source of knowledge and a unique one, appreciating its quirky yet straightforward idea
+
+### Strengths
+**Leveraging Ideas from Other Papers for Enhanced Parameter Efficiency:** This paper skillfully incorporates concepts from existing research to optimize parameter efficiency.
+
+**Achieving Comparable or Superior Performance to NOLA:** This research attains performance levels akin to LoRA while significantly enhancing parameter efficiency.
+
+### Weaknesses
+**Poorly presented results**-  The main issue in the presentation of the results lies in their lack of clarity and explanatory depth. Firstly, the results fail to offer any substantial insights into how the method operates, leaving readers without a clear understanding of the underlying mechanisms. Additionally, Tables 1 and 5 are presented as mere lists of numbers without the necessary context or explanation, making it challenging for the audience to derive meaningful conclusions from the data. A critical element that appears to be missing is a discussion of what works better and the reasons behind it, which is crucial for a comprehensive understanding of the findings. To improve the presentation of the main results, it is essential to provide better explanations and context for the data, as well as a deeper analysis of what drives the observed outcomes.
+
+### Questions
+1. GPT2-L and GPT-2M seems to perform the same for LoRA. Is there any explanation on why this is the case?
+2. The presentation of results preceding the training details in Section 3.1 seems to be an inadvertent oversight. To enhance the logical flow of the content, it is advisable to reverse the order, placing the training details before the results.What happens when you increase the number of parameters for NOLA?  - Does it perform better than LoRA. For example results of NOLA with 0.35M parameters
+3. How does NOLA's performance change when the number of parameters is increased? Does it outperform LoRA? For instance, are there any results available for NOLA with 0.35 million parameters?
+
+### Soundness
+3 good
+
+### Presentation
+2 fair
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 2
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+The paper proposes a new approach for fine tuning LLMs for downstream tasks. The key idea is to replace the low rank updates of LoRA with linear combinations of fixed random matrices for which only the coefficients need to be tuned and stored in memory which significantly reduces the storage cost. The authors present experiments in both language and vision tasks where their approach preserve the accuracy of LoRA while reducing the parameter count by half or more.
+
+### Strengths
+1. The authors propose a novel, intuitive, and principled approach to address the problem of task based fine tuning of transformer based models.
+
+2. The proposed approach shows significant reduction in storage overhead without compromising on accuracy across a range of experiments in both language and vision tasks.
+
+### Weaknesses
+1. The technical novelty is relatively minor with the overall idea being a combination of prior works PRANC and NOLA. While this seems enough to provide empirical improvement, the approach itself is not that big of an innovation over prior works. 
+
+2. While the prior approach PRANC is directly modified by the authors in this work there are no direct comparisons with it in either the language or vision tasks used to evaluate the proposed approach. There is a comparison of training loss in Section 3.4 and a comparison of the rank of possible solutions of the two approaches in Section 3.5 but without a direct comparison of test accuracy it is unclear if this approach is indeed an improvement over the baseline that it directly modifies.
+
+### Questions
+1. Why is the training time of NoLA with shared random basis similar to that of LoRA when the training time of NOLA with a unique random basis is higher? Aren't the number of coefficients being trained, the same in both cases?
+
+2. The ablation study at the end of Section 3.1 appears inconclusive. Is there any takeaway on the effect of varying the rank in NOLA?
+
+3. In Section 3.2 if only $\alpha$ and $\beta$ are quantized while A and B are not then won't that be less memory efficient than quantization in LORA?
+
+4. Please highlight the entries in Table 5 with the best performance for a given scenario. Currently there are too many entries, and it is too difficult to figure out which method is better for which case.
+
+5. If each matrix in PRANC has size $d^2$ then why do we need multiple matrices to cover the rank of the original $\Delta W$ matrix (which also has size $d^2$)?
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 3
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+The paper looked at the problem of memory requirements for Low-Rank Adaptation (LoRA) and proposed NOLA to break the rank one lower bound present in LoRA. The core concept behind NOLA is to reparameterize a neural network using a linear combination of pseudo-randomly generated weights.
+
+Thanks to the authors for a more detailed explanation of the motivations for the paper and for some of the latest research supporting them. Therefore, I will increase the rating by 1 point.
+
+### Strengths
+1. The paper discusses related works in detail and clearly summarizes its own contributions.
+2. The paper performs extensive experiments to compare NOLA and existing PEFT solutions.
+3. NOLA decouples trainable parameters from the choice of rank and the network architecture.
+
+### Weaknesses
+1. The work may need more rationales upfront to motivate the problems (i.e. the rank one lower bound present in LoRA). Given that mainstream GPUs have tens of GB of memory, it is reasonable to reduce the memory requirements from tens of GB to tens of MB at the expense of model quality through LoRA, as this can indeed reduce resource consumption and greatly reduce LLM transition overhead during inference. However, I don't think it makes much sense to further reduce memory requirements to several MBs at the expense of model quality.
+
+### Questions
+If users want to use the trained model on different versions of GPUs or software, how to ensure the consistency of the trained model? In such a situation, the same seeds can not generate the same pseudo-random matrices.
+
+### Soundness
+2 fair
+
+### Presentation
+3 good
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 4
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper proposes a technique to reduce the number of trainable parameters while fine-tuning large language models and large vision encoders. Their technique involves modeling the weight update matrix as a linear combination of fixed random matrices that are also rank-constrained. The fine-tuning process then involves learning just the coefficients in the linear combination. When the proposed technique is used, any updates to the models requires the communication/ storage of only the coefficients and the random seeds required to generate the codebook matrices (apart from the base weights of the original model of course). 
+
+The authors demonstrate that the proposed technique preserves the performance on the fine-tuning task while achieving large ( up to 1/20 x baseline models) reduction in the number of trainable parameters. This is shown in both language and vision domains. In the language domain, they show it on the NLG challenge dataset. In the vision domain, they show is on CIFAR, CUB and Caltech-101 datasets.
+
+### Strengths
+- The paper is well motivated and clearly written. 
+- Parameter-efficient fine-tuning is a popular area of research currently and this paper makes a good contribution to this area.
+- The proposed technique is interesting and the results demonstrate that the method preserves performance while achieving a low parameter count 
+- The proposed method helps overcome the some of the limitations of methods such as LoRA, as described in the paper
+
+### Weaknesses
+I appreciate the results provided in the paper. But I think that some more in-depth evaluation and some more explanation of the current results would add value to the paper. I outline some specifics below.
+
+Language experiments: 
+- Why does full fine-tuning achieve much lower performance in Table 1?
+- Could the authors also provide the performance of the GPT-M, L models *without any fine-tuning* on the tasks considered? This will give the readers an idea of how much improvement is being achieved. 
+
+Vision experiments: I find the experiments provided in this section (3.3) a bit weak. The main reason being that the ViT models pre-trianed on imagenet are already pretty powerful. Fine-tuning these models on much smaller and easier datasets such as CIFAR may not be the best way to demonstrate usefulness. In particular, I have the following comments:
+
+1. Can the authors try their techniques on more challenging datasets? 
+2. Although the number of parameters for the linear layer baseline depends on the dataset, it would be good to have this information visible in the Table. 
+3. As before, can the authors provide the performance of the models considered without any fine-tuning? (0-shot classification on the downstream datasets)
+
+### Questions
+1. What is the distribution used to generate the random basis matrices? Did the authors experiments with a few different choices?
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good

human_reviews/U2ZtvonVQz.md

@@ -0,0 +1,220 @@
+# DC-PINNs: Physics-Informed Neural Networks for Solving Derivative-Constrained PDEs
+
+- Decision: Reject
+- Scores: 3, 3, 3, 3
+
+## Abstract
+Physics-Informed Neural Networks (PINNs) have emerged as a promising approach for solving partial differential equations (PDEs) using deep learning. However, standard PINNs do not address the problem of constrained PDEs, where the solution must satisfy additional equality or inequality constraints beyond the governing equations. In this paper, we introduce Derivative-Constrained PINNs (DC-PINNs), a novel framework that seamlessly incorporates constraint information into the PINNs training process. DC-PINNs employ a constraint-aware loss function that penalizes constraint violations while simultaneously minimizing the PDE residual. Key components include self-adaptive loss balancing techniques that automatically tune the relative weighting of each term, enhancing training stability, and the use of automatic differentiation to efficiently compute derivatives. This study demonstrates the effectiveness of DC-PINNs on several benchmark problems, from basic to complex, such as quantitative finance and applied physics, including heat diffusion, volatility surface calibration, and incompressible flow dynamics. The results showcase improvements in generating solutions that satisfy the constraints compared to baseline PINNs methods. The DC-PINNs framework opens up new possibilities for solving constrained PDEs in multi-objective optimization problems.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+This paper introduces a variation on PINN that adds derivative constraints to ensure accuracy of MLP approximation of the solution to a given PDE. This variation introduces the ReLU of the constraint residuals as the additional loss term together with the three vanilla PINN losses. To balance these loss terms, this paper also proposes weights for all four losses, and an adaptive algorithm to adjust these weights during training. The new approach is tested on three types of PDE: heat, Black-Scholes and 2D Navier-Stokes.
+
+### Strengths
+The proposed method for balancing the load is presented in a logical way and seems  to improve the previous balancing algorithms by Wang, et al.
+
+The method is implemented on three sets of benchmark problems.
+
+### Weaknesses
+(1) It is unclear how realistic the problem set up is. For a genuine PDE solver with initial and boundary data, derivative information within the domain is typically unknown. Incorporating this information into the loss function introduces data leakage, as it relies on knowledge from output quantity of interest. Here is a detailed argument about the meaning of having this extra information.
+
+Take the 1D heat equation in the paper as an example, if we change the initial data to $ u(t=0, x) =\sin(2\pi x) $.
+The derivative constraints $ \frac{\partial^2 u}{\partial x^2} \leq 0, \quad \frac{\partial u}{\partial t} \leq 0 $
+will destroy the existence of the unconstrained solution $u(t,x)=e^{-4\pi^2\lambda t}\sin(2\pi x)$ for all $t\geq0$, but the unconstrained problem itself is a well-posed PDE problem and vanilla PINN solves it at acceptable accuracy.
+
+If constraints are adjusted accordingly to allow existence of solution, it becomes a clear source of data leakage (setting correct derivative bounds to allow existence of solution).
+
+Some of the PDEs have solutions intrinsically large derivatives, so that adding derivative constraints modifies the genuine solution.
+We have an example from original PINN paper:
+$
+    u_t + u u_x - \frac{0.01}{\pi} u_{xx} =0 \qquad (t,x)\in[0,1]\times[-1,1] 
+$, $u(0, x) = - \sin(\pi x),$ and $u(t, -1) = u(t, 1) = 0.$
+Set the constraints are posed as
+$ \left| \frac{\partial u}{\partial x} \right| \leq \text{bound}, $. Then, 
+even if we notice that setting $\text{bound}<\pi$ leads to the initial condition violating the constraints so that the bound is adjusted to be larger without prior knowledge of the solution, $\frac{\partial u}{\partial x}$ will eventually reach the bound as a shockwave, so the solution from DC-PINN beyond this period will diverge from the true (unconstrained) solution or/and violate the derivative constraints.
+
+Due to the possible data leakage problem above, improved accuracy may solely source from this extra info, diminishing the effectiveness of the loss balancing mechanism as another claimed contribution of the paper
+
+(2) Presentation quality
+- Equation (1) and (2) have an undefined variable $y$. From context, it is expected to be $u$.
+
+- Equation (14) has $u(x, t)$ as the argument order ($x$ first, $t$ second) but in the description (line 298), it appears as $u(t, x)$, which are inconsistent. It is recommend to be consistent.
+
+- Figure 7 captions are wrong: none of figures mention FEM, but caption says (a) is FEM. From context, it is expected to be (a)$L_0$, (b)$\mu_1$, (c) $\mu_2$.
+
+(3) Need more comparisons:
+
+The paper mentions that the proposed load balancing algorithm is derived from the method of Wang et al 2023. It is expected to have some comparisons with the previous method. 
+
+Moreover, there are other baseline methods that can be compared with the proposed load balancing technique: 
+
+- Kim, et al, DPM: A novel training method for physics-informed neural networks in extrapolation, AAAI, 2021. 
+
+- Yao, et al, Multiadam: Parameter-wise scale-invariant optimizer for multiscale training of physics-informed neural networks, ICML, 2023.
+
+### Questions
+The main question is included in the first point of the weaknesses above.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 2
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+5
+
+### Summary
+This paper considers the problem of incorporating derivative constraints when solving PDEs using PINNs. It incorporates those constraints as additional terms into the loss function and proposes to use loss balancing techniques between not only loss terms, but also the samples within each loss. The proposed method is compared to normal PINNs for different PDEs.
+
+### Strengths
+- The proposed method is better at satisfying derivative constraints than regular PINNs (that do not incorporate any derivative information).
+- The paper proposed loss balancing methods appears to have beneficial performance effects (although testing is severely limited).
+- The structure and writing is mostly clear (despite the presence of several typos in the formal parts).
+
+### Weaknesses
+The major weaknesses of the paper are: (1) incremental contribution; (2) limited motivation; (3) scarce experimental validation; (4) weak guarantees and formalism.
+
+1. **Incremental contribution:** The method used by the work to address derivative constraints is no different than that from regular PINNs in (Raissi et al., 2019), namely, combining automatic differentiation and penalties. The penalty used is in fact typical from classical constrained optimization methods, such as augmented Lagrangian (see, e.g., Bertsekas, “Constrained Optimization and Lagrange Multiplier Methods,” 1982). In other words, no additional contribution is needed to account for such constraints.  
+Additionally, the proposed balancing method is essentially a combination of (Wang et al. 2023) and (McClenny & Braga-Neto 2020). In particular, the sample balancing adaptation (12) is the same as in (McClenny & Braga-Neto 2020), while the loss balancing uses the method proposed in (Wang et al. 2023), replacing a moving average and l2 norm by an exact average and l1 norm. Neither change is clearly motivated (see (4) below). While the combination of these two methods is novel, it is unclear why or when it is needed (see (3) below). While the manuscript acknowledges its inspiration in (Wang et al. 2023) and (McClenny & Braga-Neto 2020), it does not properly state these similarities (see (2) below). For instance, the idea of balancing losses and samples within losses already appeared, e.g., in (Wang et al. 2023).
+
+2. **Limited motivation:** the paper does not properly motivate the need for derivative constraints, either in its introduction or in the experiments. This is a particularly critical issue in view of point (1) above. For instance, the derivative constraint proposed for the heat equation appears to arise from the knowledge of the phenomenon, i.e., of the PDE solution. This is in fact stated in the text: “We can add derivative constraints [...] based on the fact that the specified derivatives of the analytical solution should be negative in the defined space.” In other cases, the derivative constraints appear to be part of the boundary conditions of the problem, e.g., in the finance and Navier-Stokes case. Comparison to regular PINNs in all these settings is therefore completely unfair (see (3) below), seen as regular PINNs either do not use the additional information (heat equation) or does not even tackle the same problem (finance/Navier-Stokes). What is more, the no-arbitrage constraints in the finance application do not appear to be of the same type considered throughout the text, since they must consider a two-sided inequality, namely (26). In the case of the Black-Scholes model, the motivation for using NNs is unclear since a closed-form solution can be obtained somewhat straightforwardly. Finally, the paper provides very limited discussion on its relation to prior work, particularly when it comes to other approaches to incorporate constraints when solving PDEs. There are a plethora of papers on incorporating boundary conditions and other constraints, such as Chalapathi et al., ICLR’24; Gao et al., J. Comp. Phys.’21; Psichogios et al., AIChE J.’92; Lu et al., 2021; Basir et al., J. Comp. Phys.’22, to name only a few. Any of these would also provide much better baselines (see (3) below).
+
+3. **Scarce experimental validation:** The only comparison to other methods in the paper is with traditional PINNs. No other methods for incorporating constraints into PINNs are considered. Regular PINNs are, however, a poor baseline since it does not even tackle the proposed PDE problems. This is particularly critical in the finance and Navier-Stokes applications, where the PDE problems considered have boundary conditions incorporating derivative inequalities (see (16-17) and (18-19)). In that sense, it is not surprising that, e.g., DC-PINNs perform better than PINNs in Figure 2 or Figure 6, since PINNs do not account for the derivative conditions in the problem formulations. Additionally, there is no experimental validation of the proposed weight balancing scheme. This is an issue particularly since it is not well-motivated (see points (1) and (4)). A better baseline would be, e.g., using PINNs with the derivative penalty term with fixed weights. This would also help understand the role and necessity of the different weighting schemes and better motivate their combination. Since no theoretical motivation or guarantee is provided (see (4) below), this lack of experimental evidence is all the more critical.
+
+4. **Weak guarantees and formalism:** While the paper is for the most part well-written, its lack of formalism makes it hard to follow at times. Firstly, several typos throughout the derivations make the presentation confusing, e.g., the use of $u$ for the PDE solution in line 78 immediately followed by calling it $y$ in (1-2) and throughout the rest of the text. Taken formally, (7-8) also do not make sense since the “inequality” to be satisfied in (8) is $x \geq 0$ and the argument of $\gamma$ in (7) is always non-negative, namely, $| h(\phi(x)) |^2$. I assume this is a typo since it is corrected in (11). What is more, the derivative constraints considered in the work are of the form $h(\phi(x)) \geq 0$. Hence, the loss $\mathcal{L}_h$ should have a negative sign (seen as $\mathcal{L}$ is being minimized and the value of $h$ must be increased in order to satisfy the constraint). These recurring typos in the formalism of the paper not only confuse the reader, but also suggest that the manuscript needs more attention before being published.  
+Another critical point is the motivation for the use of the l1 norm (absolute value) in (13) as opposed to the l2 norm (squared values) from Wang et al., 2023. The text states that “the main reason for the choice of absolute values [...] is to avoid overlooking outlier violations of the inequality constraints.” This justification does not make sense to me. In fact, the squared value gives considerably more weight to outliers than the absolute value, which is the whole reason why the latter is used in robust statistics, e.g., the Huber loss, or for using the l1 norm in compressed sensing. The manuscript goes on to justify that “most elements of $\nabla\mathcal{L}_h$ are assumed to be zero values in almost all cases.” They would therefore not need the l1 geometry to force their sparsity, in which case the squared value would once again make sense (as is the case in, e.g., elastic net).  
+Finally, it should be noted that the use of penalties with weight adaptation heuristic does not enforce the derivative constraints, but only penalizes their value. This can be seen in Figure 4, where even DC-PINNs do not fully satisfy the constraints. What is more, it could be that satisfying the derivative constraints would lead to violating the PDE or boundary conditions (this can be done by, e.g., making $\lambda_h \to \infty$). It is hard to tell from Figure 3 whether that is the case since no numerical values are provided for the errors. In fact, neither PINNs nor DC-PINNs appear to fit the ground truth well in this case. In that sense, it would also be more informative if Figure 4 showcased the value of the constraints so as to show how much it is violated, i.e., having a continuous rather than binary heatmap. This would be in line with the rationale provided in the text on line 143 stating that “setting large loss weights can cause an ill-conditioned problem. On the other hand, when small loss weights are chosen, the estimated solution may violate the inequalities.” As it is, this statement is not justified, either theoretically or empirically. In particular, without any theoretical guarantee or substantial experimental validation (see (3) above) of the proposed methods, it is hard to justify the “derivative-constrained” name.
+
+### Questions
+See Weaknesses. Here are additional minor points:
+
+1. How are the different terms in the PINN baseline weighted? This is quite standard and known to have big impacts on performance. How does the DC-PINN compare to PINN including the derivative penalty with a fixed weight chosen by, e.g., cross-validation?
+1. How is the constraint (26) handled in the finance application? In other words, how is the two-sided inequality handled when the penalty in (7) is designed for $D(y(x)) \geq 0$?
+1. On line 453, it says “Figure 6 compares the pressure fields obtained from models at a representative time …”. What is meant by a representative time step? It would be good to have some performance metric aggregated for all times.
+1. In the caption of Figure 5, should “verocity��� read “velocity”?
+1. In general, the labels and legends of the figures are too small (especially the legends).
+1. The bibliography should be updated to reflect the current status of papers. E.g., (McClenny & Braga-Neto 2020) has already been published for over a year.
+
+In view of these issues, I do not recomment that the paper be published in its current state.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+1
+
+---
+
+## Human Reviewer 3
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+This paper develops a Derivative-Constrained PINNs framework to incorporate additional inequality constraints, beyond the typical PDE, boundary, and initial condition constraints, into the optimization process of PINNs. The primary contribution is the introduction of a self-adaptive loss balancing technique that automatically adjusts the relative weighting of various loss penalty terms, as well as the individual weighting within each type of penalty term. The authors demonstrate the effectiveness of their proposed method through several practical problems, highlighting DC-PINNs' ability to recover PDE solutions while satisfying specified inequality conditions on derivatives.
+
+### Strengths
+The paper is well-organized and clearly written, making it easy to follow overall. It provides strong motivation for incorporating inequality derivative constraints into the optimization process. The proposed method is clearly explained and accessible, with practical and well-known benchmark problems that demonstrate its effectiveness in addressing real-world challenges.
+
+### Weaknesses
+•	Novelty: Inequality-constrained PINNs have been studied in existing work, such as Lu et al. [1]. Although Lu et al. [1] addresses inverse problems, while this work directly solves a PDE, similar problems are considered using different methods. Additionally, the benchmark method from Lu et al. [1] should also be included for comparison.
+
+•	Insufficient Experimentation: For example, in the heat equation experiment, the heat equation coefficient $\lambda$ is set to 0.1. However, PINNs have well-known failure modes: when PDE coefficients are small, PINNs can recover the PDE solution, but they fail when PDE coefficients are large, as noted by Krishnapriyan et al. [2]. The proposed method should demonstrate robustness across a range of coefficient values. Additional, more robust experiments to hyperparameters such as NN size would be good.
+
+[1] Lu, Lu, et al. "Physics-informed neural networks with hard constraints for inverse design." SIAM Journal on Scientific Computing 43.6 (2021): B1105-B1132.
+
+[2] Krishnapriyan, Aditi, et al. "Characterizing possible failure modes in physics-informed neural networks." Advances in neural information processing systems 34 (2021): 26548-26560.
+
+### Questions
+In the traditional penalty method, the penalty coefficient is required to go to infinity for the transformed unconstrained problem to converge to the original constrained problem. Here, the method you are using is still a penalty method. Nevertheless, the penalty coefficients $\lambda_{\beta}$ or $m_{\beta}^{i}$ in your work may not keep increasing if $\nabla \hat L_{\beta} (k)$ is small, even if the constraint violation remains large. I wonder if this could occur due to a complex loss landscape or potential ill-conditioning issues such as when the PDE coefficient is large.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 4
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+The paper extends PINNs to derivative constrained PDEs by introducing a term in the PDE loss. Specifically, the authors target solving PDE with inequalities on the derivative constraining the solution. The proposed method is then validated on several examples.
+
+### Strengths
+The paper is well-written and easy-to-follow. The contribution is simple to understand and to my knowledge, no work have studied thus kind of PDE solving problem with derivatives constraints. 
+Several examples are shown to illustrate the method. 
+The example chosen to evaluate are relevant and taken from several literature. 
+The method is well-described and experiment settings are detailed.
+
+### Weaknesses
+The contribution consists in the and to adding a term in the PINNs loss to model the addition of the constraints as already made in the standard PINNs proposed by Raissi et al. These constraints can be linked to Boundary condition problems, which is a well-known context of PINNs. In such context, the derivative constraints are treated the same as in the proposed method (initial and boundary residuals). The loss balancing algorithm are taken from existing works. 
+In the experiments, if other solution can be found by PINNs by not considering the added constraints, doesn’t it mean that the problem is ill-posed? Meaning that several solutions exist for the given PDE. This is unfair in my opinion to compare the proposed algorithm with method which (i) do not have any information about the added constraint (ii) have to solve ill-posed problems. 
+Doesn’t any baseline exist to solve this type of PDE-constrained problems that could be added as a comparison?
+
+### Questions
+- Are samples used during training? If so, where are they located? How many of them are used? 
+- Why aren’t MLP results displayed and commented? 
+- How does the different models behave when adding too much constraints? i.e. when the system does not admit any solution? 
+- In the example 6.3, why the settings for PINNs and DC-PINNs are not the same? (last paragraph, velocity and pressure fields) What are the results when aligning settings? 
+- In example 6.3, It seems that PINNs successfully learns the parameters of the PDEs? Could the author elaborate on the relative performances of PINNs wrt DC-PINNs? What are the advantages of DC-PINNs in this context? 
+- In example 6.4, it seems that DC-PINNs is much longer to train. Could you elaborate about this aspect? Why does the training time increases this much? 
+
+Additional remarks:
+-	Typo line 249 with “Eq.equation 9”
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+2

human_reviews/WcZLG8XxhD.md

@@ -0,0 +1,178 @@
+# Learning-Augmented Frequent Directions
+
+- Decision: Accept
+- Scores: 8, 6, 8, 8
+
+## Abstract
+An influential paper of Hsu et al. (ICLR'19) introduced the study of learning-augmented streaming algorithms in the context of frequency estimation. A fundamental problem in the streaming literature, the goal of frequency estimation is to approximate the number of occurrences of items appearing in a long stream of data using only a small amount of memory. Hsu et al. develop a natural framework to combine the worst-case guarantees of popular solutions such as CountMin and CountSketch with learned predictions of high frequency elements. They demonstrate that learning the underlying structure of data can be used to yield better streaming algorithms, both in theory and practice.
+
+We simplify and generalize past work on learning-augmented frequency estimation. Our first contribution is a learning-augmented variant of the Misra-Gries algorithm which improves upon the error of learned CountMin and learned CountSketch and achieves the state-of-the-art performance of randomized algorithms (Aamand et al., NeurIPS'23) with a simpler, deterministic algorithm. Our second contribution is to adapt learning-augmentation to a high-dimensional generalization of frequency estimation corresponding to finding important directions (top singular vectors) of a matrix given its rows one-by-one in a stream. We analyze a learning-augmented variant of the Frequent Directions algorithm, extending the theoretical and empirical understanding of learned predictions to matrix streaming.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4
+
+### Summary
+The paper discusses the problem of online learning of frequent directions given a stream of vectors. In particular, they also consider the learning-augmented problem, where one has access to a predictor that provides estimate of which directions are frequent. They also address the problem of frequent items, since this a special case of the frequent direction problem, where item i is associated to the standard vector e_i.
+
+The contribution of the author is two-fold. They first study the problem of frequent directions using power-law distributed data (i.e., in this case the singular values of the input matrix A, that is given online row by row, follow a power law or Zipfian distribution). To this goal, they analyze the Frequent Direction Algorithms (Ghashami et al., 2016) in this setting (Theorem 3.3). Additionally, they also shows that if one has access to a perfect learning oracle, the expected error of the algorithm can decrease. As a special case, they also shows that a variation of this algorithm (based on the Misra-Gries algorithm, that is the 1-dimensional variant of the frequent direction algorithm), allows to solve the (learning-augmented) frequent item problem with Zipfian distribution law of the frequency. In particular, they retrieve the results of the previous work with a simple deterministic algorithm, compared to previous work that is based on CountMin sketch and is randomized.
+
+### Strengths
+The paper is within the algorithms with prediction framework, that was subject to a considerable amount of research in the last years. The paper studies a new problem in the algorithms with prediction framework: learning-augmented frequent directions. 
+
+One of the main strength of the algorithm is that they are also able to recover, with a simple algorithm, results of the previous work for the special case of frequent items.
+
+Another strength of the paper is that it is extremely well-written. The technical section is easy to follow. The results look correct.
+
+Due to the simplicity of the algorithm, clarity of exposition, and the relevance of the results, I believe this paper provides a good contribution.
+
+### Weaknesses
+The paragraph on robustness 361-366 is not clear. It seems that they constraint the oracle to provide still a lot of valuable information. I believe it would be better to provide instead sufficient conditions required on the oracle to still obtain the approximation guarantees of Thm 3.4 (or resp., Thm 3.2). It seems to me that appendix E shows that one can accomodate an arbitrary bad oracle (Theorem E.1) using a modification of the algorithm. Unless I am missing something, I believe this is more important than the robustness paragraph 361-366, and should be maybe included in the main paper.
+
+Pages 15-28 contains a sequence of plots with virtually almost no-text (Figure 53 seems a very high number of figures). I think it would be better to select some plots that are then discussed with text to express a message, or summarize the information in a different and more concise manner.
+
+### Questions
+See weakness.
+
+The lower bound of Thm.1  and Thm.3 seems specific to the algorithm that is used. Is there any lower bound that applies for any algorithm, given an input sequence that satisfies the Zipfian law assumption?
+
+### Soundness
+4
+
+### Presentation
+3
+
+### Contribution
+3
+
+---
+
+## Human Reviewer 2
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+3
+
+### Summary
+This paper presents a theoretical and empirical study of learning-augmented streaming algorithms, focusing on frequency estimation and its extension to matrix streaming. The authors introduce a learning-augmented variant of the Frequent Directions algorithm and demonstrate both theoretical improvements and practical benefits when combining learned predictions with traditional streaming algorithms.
+
+### Strengths
+1. The paper provides a clear theoretical analysis of both learning-augmented Misra-Gries and Frequent Directions algorithms. Achieves state-of-the-art bounds for learning-augmented frequency estimation with a simpler, deterministic algorithm
+
+2.  Successfully extends learning-augmented techniques from frequency estimation to matrix streaming
+
+3. Comprehensive experimental evaluation on real-world datasets. Shows 1-2 orders of magnitude improvement over baseline algorithms and very close to SVD algorithm which is memory intensive.
+
+### Weaknesses
+1. Limited diversity in the prediction models used. More comparison with other learning-augmented streaming algorithms would strengthen the evaluation
+
+
+2. Could benefit from ablation studies on prediction quality vs. performance
+
+### Questions
+1. Is there a fundamental trade-off between space usage and prediction accuracy?
+2. How does the performance vary with different stream lengths?
+3. How does the approach scale to very high-dimensional data?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+3
+
+---
+
+## Human Reviewer 3
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+3
+
+### Summary
+This paper studied the learning-augmented frequent directions, which is a natural generalization of the frequency estimation problem in a stream of elements. The frequency estimation problem asks the following question: we are given a stream of $n$ items that below to at most $d$ distinct elements, what is the additive error we could get for the estimation with a limited memory $m\ll \min(n,d)$. Learning-augmented frequency estimation, which additionally assumes a (potentially erroneous) oracle that returns a predicted frequency of each element *and* the Zipfian distribution of the elements, has been extensively studied in recent years. The state-of-the-art error bound was given by ACNSV [NeurIPS’23] for frequency estimation.
+
+This paper asked if we want to generalize the frequency estimation in high-dimensional applications. Here, a natural setup is to have multiple $d$-dimensional vectors $A_1$, $\cdots$, $A_n$. Furthermore, the following aspects are defined for the generalization.
+- To generalize the notion of frequency estimation error, the paper used a notion of the error of vector projection to the singular vectors of the matrix formed by $A=[A^T_1, A^T_2, \cdots, A^T_n]^T$. 
+- To generalize the distribution assumption of Zipfian distributions, the paper used the notion of the square of the $i$-th largest singular values of the matrix $A$ being inversely proportional to $i$.
+- To generalize the frequency prediction, the paper introduced an oracle that directly predicts orthogonal frequency directions, which are the ``right rows’’ an optimal algorithm should have used.
+
+The paper showed that (learning-augmented) frequency estimation can be reduced to the (learning-augmented) frequent direction problem defined above. This justifies the claim that this problem is a *natural* extension of the frequency estimation counterpart.
+
+The paper obtained a frequent direction algorithm with $O(\log{m}\cdot \frac{\log{d/m}}{\log^{2}{d}}\cdot \frac{\||A\||^2_{F}}{m})$ additive error for $m$-memory algorithms even without predictions (by taking advantage of the ‘Zipfian’ distribution of the matrix).  For algorithms with predictions, the error bound becomes $O( \frac{1}{\log^{2}{d}}\cdot \frac{\||A\||^2_{F}}{m})$, assuming the prediction is at least accurate on a constant fraction of the predicted frequency directions. The paper further generalized their results to the frequency estimation problem, showing that it is possible to recover the best bound for the frequency estimation problems using the new algorithm.
+
+### Strengths
+My overall opinion of this paper is quite positive: it is well-motivated and well-written, and it contains quite some results that could be interesting to a wide range of audiences both in machine learning and algorithms in general. Barring some lower-order exposition issues, the paper is generally easy to follow, and the technical idea, while simple, is quite neat. As such, the paper is a clear acceptance from my perspective.
+
+### Weaknesses
+I do not see any major weakness in the paper. Some comments for expositions:
+- The column indexing of the SVD in algorithm 1 is not as clear as it could be. I suggest having a notation section that lists all the notations you use in the paper.
+- Calling the lemma statement of Liberty [Arxiv’22] a ‘fact’ is perhaps overly trivializing the significance of the lemma. I would say to call it a ‘proposition’ (I know it’s a one-line proof, but still :)).
+- The discussion about the robustness in the main paper is perhaps too sketchy. I understand that you have a more detailed discussion in the appendix; however, I think it’d be helpful if you could expand the discussion.
+
+### Questions
+In the experiment, you used SVD as a benchmark. Can you give the error bound for storing all vectors in $A$ and run SVD? In particular, does the error also scale with $\||A\||^2_{F}$?
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+4
+
+---
+
+## Human Reviewer 4
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4
+
+### Summary
+This work revisits the problem of learning-augmented frequency estimation and provides simplified, improved, and generalized results. In the classic frequency estimation problem, we wish to estimate the number of occurrences of an item in a stream up to some small error. In the “learning-augmented” setting, we assume that some oracle tells us the most frequent items, which allows for better algorithms. Furthermore, the stream is also assumed to follow a Zipfian distribution, that is, there is a clear choice of “most frequent items” so this assumption is actually useful. Prior work on this subject, such as the work of Hsu et al. (2019), uses randomized algorithms based on CountMin, CountSketch, and related constructions. The current work shows that the Misra-Gries algorithm, a classic deterministic algorithm for frequency estimation, actually obtains sharp guarantees for this problem. Furthermore, this work also generalizes this to the vector setting, where the stream items are vectors and the desired output is estimates to the singular vectors of the stream viewed as a matrix. In this setting, Misra-Gries corresponds to the frequent directions algorithm, and the authors obtain similar guarantees.
+
+### Strengths
+The paper is remarkably well written. Every time I wanted to ask a question to the authors, I felt like it was almost immediately answered in the following text. The contribution is also solid, and gives a clean improvement to prior work on this subject. The experiments were also insightful, and showed how an initial estimate of the top singular values can be used as the “learned frequent directions” for use in the learned frequent directions algorithm. In hindsight, this seems pretty similar to related ideas in numerical linear algebra about using a coarse estimate of the top singular components when designing algorithms for matrix estimation tasks, such as https://arxiv.org/abs/2010.09649 and https://arxiv.org/abs/1511.07263. Connections to such works might be interesting to include.
+
+### Weaknesses
+The only complaint is that the paper is pretty much a combination of existing and known ideas. However, it is a high quality implementation of such a study and adds valuable information to this literature.
+
+### Questions
+None at this time.
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+3

human_reviews/WxqWuG431g.md

@@ -0,0 +1,226 @@
+# The Geometry of Concepts: Sparse Autoencoder Feature Structure
+
+- Decision: Reject
+- Scores: 3, 3, 3, 3, 1
+
+## Abstract
+Sparse autoencoders have recently produced dictionaries of high-dimensional vectors corresponding to the universe of concepts represented by large language models.
+We find that this concept universe has interesting structure at three levels:
+1) The ``atomic" small-scale structure contains ``crystals" whose faces are parallelograms or trapezoids, generalizing well-known examples such as {\it (man:woman::king:queen)}.
+We find that the quality of such parallelograms and associated function vectors improves greatly when projecting out global distractor directions such as word length, which is efficiently done with linear discriminant analysis.
+2) The ``brain" intermediate-scale structure has significant spatial modularity; for example, math and code features form a ``lobe" akin to functional lobes seen in neural fMRI images. We quantify the spatial locality of these lobes with multiple metrics and find that clusters of co-occurring features, at coarse enough scale, also cluster together spatially far more than one would expect if feature geometry were random.
+3) The ``galaxy" scale large-scale structure of the feature point cloud is not isotropic, but instead has a power law of eigenvalues with steepest slope in middle layers. 
+We also quantify how the clustering entropy depends on the layer.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+3
+
+### Summary
+The manuscript claims to study the structure of features (point clouds) obtained using sparse autoencoders on some layers of the LLM Gemma. These features can be interpreted as concepts according to the literature cited. The structure is studied at three different scales, from fine to coarse: pairwise discriminability of concepts (with invariance to irrelevant distractors); clustering of concepts (comparing clustering by geometry alone and functional clustering); and global shape of the point cloud. Clustering and global shape are compared against simple null hypotheses (uniform clusters with no spatial relations, and isotropic gaussian blob, respectively).
+
+### Strengths
+Evocative language that promises (but then fails to) to explore an interesting connection across domains of knowledge.
+
+The approach to compare the similarity of geometric and functional clusters is interesting.
+
+There is value to performing a purely exploratory study of the feature space of a LLM, this might be an interesting starting point for the authors to identify a direction worth pursuing in a rigorous manner in the future.
+
+### Weaknesses
+The manuscript is written casually and in a rush. Methods are explained poorly. Evocative words from disparate fields are thrown together with no in-depth argument for why the analogies are meaningful. 
+Action: I suggest writing a Methods or Approach section, before presenting results and interpretations.
+
+The manuscript fails to convey how novel and unexpected the findings are, and what is the contribution. In particular, are the uniform clustering and the spherical gaussian valid even as a strawman? Is there some advance with respect to other literature? 
+
+Just as the writing is casual (“wreaks havoc”, “fire together” used with no explanation and then defined two paragraphs later; figure 4 explained before figure 3 before figure 2, and figure 2 caption using concepts defined in fig 3 without a pointer; many typos), so is the analysis. 
+Action: figures should be ordered according to when they are referred to in the text. Nex concepts and jargon (e.g. 'fire together') should be defined on first use.
+
+As stated above, I believe there’s value in exploratory analysis, but for a manuscript to have value to the readers perhaps there should be some in-depth analysis. But as an outsider to this field, I don’t know the standards and I might be unaware that a quick first-pass of the data is worth publishing.
+
+I am not a chemist or cosmologist, so I have a hard time judging if the analogies to crystals and galaxies are more than a very loose analogy (the brain lobe analogy is quite superficial though: “some principal axes slightly wider than others akin to a human brain”… with literally no additional articulation of this particular proposal, a watermelon would be as good a metaphor). If one could actually learn something about LLMs from leveraging the maths and physics of crystals and galaxies, that would be valuable. But this manuscript is quite far from that, in applying basic first-pass data analytics tools that are not specific to any field.
+Action: please provide more detailed explanations of how these analogies relate to the LLM features, or justify why these particular analogies were chosen over others.
+
+### Questions
+What are the significant results? What is novel? 
+
+Why are unform clustering and spherical gaussian meaningful/plausible alternative outcomes? 
+Action: please explain the rationale behind choosing these specific null hypotheses and how they relate to current understanding or assumptions about LLM feature spaces.
+
+Why is the power law “surprising” in the shape analysis, in what way is it not expected?
+Action: please provide context on what distribution or behavior they expected to see and why, which would help readers understand the significance of the power law finding.
+
+The five co-occurrence measures “all discover the code math lobe”. What about the other clusters? Are the differences between measures informative? Or worrying?
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 2
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+3
+
+### Summary
+The contribution concerns the structure of latent spaces in LLMs seen through the lense of large sparse autoencoders. This is a field of considerable interest and here interesting hypotheses are explore, including analysis at multiple scales from local to global structure. 
+
+All analyses are empirical and based on the GemmaScope open source SAEs. 
+
+ At the local scale generalized functional relations are found with an interesting de-confounder projection to discover relations as "king-man"="queen-woman" ="crowning”). At the intermediate scale are found distributed areas corresponding to topics, while interesting asymmetries are found at the global scale.
+
+### Strengths
+Several contributions are presented including projecting confounders away at the local scale to recover functional descriptions. At the intermediate level t-SNE plots are used to develop topical distributions. At the global scale a null hypothesis of isotropic covariance is rejected, to discover unexpected structure in the eigenvalues (viz the fractal cucumber mentioned in L361)
+
+### Weaknesses
+While intriguing, the whole set of analyses appear quite basic and findings appear anecdotal. Unclear how the different dimensions of the Gemma Scope data enter the analyses.
+
+ All in all the investigations could be better motivated. Would be interesting to link the discussion with human alignment and known representational relations in human culture.
+
+The terms used to denote functional scales: "atom", "brain" and "galaxy", could be better motivated.
+
+### Questions
+Make the novelty more clear - e.g., relative to already published work on SAEs and Gemma Scope in particular (viz. the work using UMAP rather than t-SNE for visualization of topics). 
+
+Consider discussing the structure found in the context of the work reported in "related work"
+
+Would there be a away to understand the similarities and differences in Fig 5?
+
+How robust are the results in Fig 7 wrt different latent space dimensions of the SAE?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 3
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+2
+
+### Summary
+This paper explores the geometry of structures in the relationship between concepts as revealed by sparse autoencoder in LLM at different scales: "crystals" at the atomic scale, spatially localized modules at the brain scale, and anisotropy of the feature point clouds in the so-called galaxy scale.   Factorizing out the distractor dimensions using LDA has been shown to be helpful in revealing concept clusters better. The co-occurring statistics between pairs of SAE features reveal spatialized localized functional modules. On a large scale, the SAE point cloud is found to be strongly clustered and follows power laws.
+
+### Strengths
+The paper systematically studied some geometric properties in the relationship among concepts in LLM model at different scales, and made some interesting observations.
+
+### Weaknesses
+However, the computational and functional implications of these observations on the geometry of concepts relationship at different scales need to be clarified.   Why are these observations interesting? Why do they matter? What are their utilities? How do they help us understand or improve LLM? It would be helpful to address these basic questions in the introduction and discussion.
+
+### Questions
+Why are these observations interesting? Why do they matter? What are their utilities? How do they help us understand or improve LLM? It would be helpful to address these basic questions in the introduction and discussion.
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 4
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+3
+
+### Summary
+This paper examines multi-scale structures in sparse autoencoder (SAE) feature spaces within language models. It proposes that SAE features exhibit patterns at three levels: atomic (e.g., crystal shapes like parallelograms), brain-scale (modularity similar to brain lobes), and galaxy-scale (clustering with a power-law eigenvalue distribution). The study applies clustering, dimensionality reduction, and co-occurrence measures to suggest a structured feature space, potentially advancing model interpretability.
+
+### Strengths
+- The multi-scale structural analysis is ambitious, with an interesting attempt to parallel functional modularity in neural networks to concepts from biological brains.
+- The methodological variety is a strength, as the paper employs clustering and linear discriminant analysis to explore SAE feature structure at multiple levels.
+- Visualizations of proposed structures (e.g., crystal formations and modular lobes) aid in explaining complex ideas, though they lack robust validation.
+
+### Weaknesses
+- The writing is inconsistent, with rough sections, minor errors (e.g., “Instand” on line 407), and inconsistent citation formats, suggesting a need for refinement. Additionally, the paper’s presentation lacks depth in essential sections, more closely resembling an extended abstract than a full conference paper.
+- The paper does not make full use of the standard length for conference papers, leaving room to expand the introduction, clarify methodology, and provide further experimental detail. For example, the introduction lacks comprehensive background or positioning within existing literature, and the methodology could be detailed to improve clarity. Additionally, there is no supplementary code provided, which limits reproducibility and replicability.
+- The identified patterns and modular "lobes" have limited novelty, as these structural insights are well-documented in language model literature. The study does not significantly advance the current understanding of SAE feature structures or provide actionable interpretability benefits.
+- Claims of crystal structures and functional lobes are interesting but lack rigorous validation. The parallelogram and trapezoid analogies remain speculative without robust quantitative analysis, making it unclear whether these patterns consistently represent SAE structures or are isolated findings.
+- The clustering methods are not sufficiently justified or compared with alternatives that might yield more interpretable groupings, such as unsupervised methods better suited to high-dimensional data.
+
+### Questions
+See above sections for details.
+
+### Soundness
+1
+
+### Presentation
+2
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 5
+
+### Rating
+1
+
+### Rating Number
+1
+
+### Confidence
+5
+
+### Summary
+This paper analyzes sparse autoencoders and provides an interpretation of the structure that they learn. They provide analogies to the low, mid and high level in terms of solid state physics, neuroscience, and astronomy respectively.
+
+### Strengths
+I believe these comparisons are certainly original and that this paper has the potential to help understand a specific but common architecture for deep learning. Some of the figures look nice.
+
+### Weaknesses
+This should have been submitted in the next cycle rather than rushed for this one. This was not reviewed by the authors for clarity, with issues ranging from minor (first entry of bibliogrophy an erroneous duplicate of the second) to the critical (the introduction + related work combined are 3 paragraphs). These sections are so spartan that I am unable ascertain what the authors were necessarily trying to accomplish, much less whether they succeeded in their goal. The conclusion does not enlighten at all.
+
+### Questions
+None
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1

human_reviews/X7gqOBG8ow.md

@@ -0,0 +1,301 @@
+# Generalizing Denoising to Non-Equilibrium Structures Improves Equivariant Force Fields
+
+- Decision: Reject
+- Scores: 5, 5, 5, 5, 6
+
+## Abstract
+Understanding the interactions of atoms such as forces in 3D atomistic systems is fundamental to many applications like molecular dynamics and catalyst design. 
+However, simulating these interactions requires compute-intensive \textit{ab initio} calculations and thus results in limited data for training neural networks. 
+In this paper, we propose to use \textbf{\underline{de}}noising  \textbf{\underline{n}}on-equilibrium \textbf{\underline{s}}tructures (\textbf{DeNS}) as an auxiliary task to better leverage training data and improve performance. 
+For training with DeNS, we first corrupt a 3D structure by adding noise to its 3D coordinates and then predict the noise. 
+Different from previous works on denoising, which are limited to equilibrium structures, the proposed method generalizes denoising to a much larger set of non-equilibrium structures.
+The main difference is that a non-equilibrium structure does not correspond to local energy minima and has non-zero forces, and therefore it can have many possible atomic positions compared to an equilibrium structure.
+This makes denoising non-equilibrium structures an ill-posed problem since the target of denoising is not uniquely defined.
+Our key insight is to additionally encode the forces of the original non-equilibrium structure to specify which non-equilibrium structure we are denoising.
+Concretely, given a corrupted non-equilibrium structure and the forces of the original one, we predict the non-equilibrium structure satisfying the input forces instead of any arbitrary structures.
+Since DeNS requires encoding forces, DeNS favors equivariant networks, which can easily incorporate forces and other higher-order tensors in node embeddings.
+We study the effectiveness of training equivariant networks with DeNS on OC20, OC22 and MD17 datasets and demonstrate that DeNS can achieve new state-of-the-art results on OC20 and OC22 and significantly improve training efficiency on MD17.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper introduces DeNS, an approach to improve energy and force predictions with the aid of non-equilibrium structures' denoising as an auxiliary task. Its implementation feeds forces from original structures as inputs, contributing to a well-structured problem. Demonstrated results indicate minor improvements on EquiformerV2 for datasets like OC20, OC22, and MD17.
+
+### Strengths
+-The paper addresses an essential challenge: the development of a self-supervised learning methodology using non-equilibrium molecules.
+
+-The proposal offers a unique perspective on non-equilibrium denoising by discussing the ill-posed mapping. The handling of input encoding using forces seems logically feasible.
+
+-The documentation is unambiguous and comprehensible.
+
+### Weaknesses
+-The paper could benefit from a broader theoretical discussion and a comparative analysis with other self-supervised techniques for non-equilibrium structures, such as denoising pretraining in [1], Noisy Nodes (using the OC20 dataset)[2], and improved noisy nodes (using MD17)[3]. These techniques have demonstrated efficacy for energy or force predictions for non-equilibrium molecules, hence their significance.
+
+-The motivation behind the paper's approach needs additional validation.
+a) The concept of “encoding force as input” finds extensive discussion in the paper. However, this approach needs corroborative proof from experiments. Results from Table 1e indicate energy prediction outcomes remain the same without force encoding. The possibility of label leakage and its contribution to improvement in force prediction, as discussed under question 2, needs examination.
+b) Similar problem-solving approaches have been published. A comparative discussion highlighting the distinctions and superiority of this paper's proposed methodology would prove advantageous.
+
+-The significant results were, to a large extent, achieved through Equiformer. Against Equiformer's backdrop, the improvements contributed by DeNS appear minimal.
+
+[1] Yuyang Wang, Chang Xu, Zijie Li, and Amir Barati Farimani. Denoise pretraining on nonequilibrium molecules for accurate and transferable neural potentials. Journal of Chemical Theory and Computation, 2023. 
+[2] Feng, S., Ni, Y., Lan, Y., Ma, Z. & Ma, W.. (2023). Fractional Denoising for 3D Molecular Pre-training. Proceedings of the 40th International Conference on Machine Learning, in Proceedings of Machine Learning Research 202:9938-9961 Available from https://proceedings.mlr.press/v202/feng23c.html.
+[3] Jonathan Godwin, Michael Schaarschmidt, Alexander L Gaunt, Alvaro Sanchez-Gonzalez, Yulia Rubanova, Petar Velickovi ˇ c, James Kirkpatrick, and Peter Battaglia. Simple GNN regularisation for 3d molecular property prediction and beyond. In International Conference on Learning Representations, 2022.
+
+### Questions
+-While encoding input forces can mitigate the non-equilibrium denoising's ill-posed problem, [1] shows that denoising without force input is also plausible. Does this undermine your motivation and imply the redundancy of input force encoding?
+-Could there be label leakage when encoding forces as input for energy and force predictions of structures?
+-Given that force prediction is one of your experiments, should you consider adding the force prediction loss to eq. (6)?
+-Provision of the pseudocode for DeNS training would be beneficial. This can offer insights into the usage of Multi-Scale Noise and other hyperparameters like p_{ DeNS }.
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 2
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+In this paper, the authors proposed a denoising non-equilibrium structures (DeNS) training strategy to improve force field learning. Different from previous denoising approaches limited to equilibrium structures, DeNS enables the utilization of non-equilibrium structures for the denoising task by encoding the corresponding non-zero forces for specifying the denoising targets. Extensive experiments are conducted to demonstrate the effectiveness of DeNS.
+
+### Strengths
+- The target problem is of interest to the machine learning force field community. 
+- The proposed approach to encode the forces for specifying target structures when denoising perturbed non-equilibrium structures is a new modification compared to previous methods like noisy-node.
+- The experimental evaluation covers both small and large-scale benchmarks. The ablation studies on the introduced hyperparameters are informative for practitioners to try the proposed DeNS approach in their applications.
+
+### Weaknesses
+- **Regarding the performance improvement** . Serving as a simple modification upon previous denoising training strategy on molecular modeling, performance improvement is the most important aspect to measure the quality of this work. However, there exist several issues that need to be further clarified:
+  - The gains brought by DeNS diminish with the dataset scales up. From Table 1 and Table 2, we can see that models trained on the 2M subset of the OC20 S2EF dataset benefit a lot more from the DeNS auxiliary task compared to the OC20 S2EF-All+MD split.
+  - The improvement on MD17 is limited compared to the OC series experiments.
+  - The improvement on the IS2RE task is also limited (sometimes DeNS even hurts the performance).
+  - The gains on the energy and force metrics are not consistent across different datasets with different scales. In the OC20 tasks, models trained with the DeNS task achieve lower force errors and competitive energy errors compared to the standard training and vice versa for the OC22 tasks.
+
+The first issue relates to the scaling property of the proposed training strategy. I recommend the authors further design experiments to investigate whether such a phenomenon is due to the inability of the proposed DeNS to bring performance gains when large-scale data is provided or other factors of the model, optimization, and so on. The second issue relates to the generality of the proposed strategy. MD17 contains simple molecules instead of adsorbate-catalyst complex in OC20/22. This dataset is much smaller than OC20/22 dataset, on which the proposed DeNS is expected to bring more gains according to the phenomenon mentioned in the first issue. However, the results in Table 5 and 6 show the gains are limited. It is suggested to further verify the generality of DeNS. The third issue relates to the significance of the force field learning task. In OC and other similar applications, what we really care about is to obtain the property of the equilibrium states like relaxed energy and structure. The error of force field model is an indirect metric. In this sense, the significance of improvement brought by DeNS on energy and force error of S2EF task should be reexamined based on the IS2RE performance.
+
+Overall, I recommend the authors to carefully clarify the proposed issues above with further experimental evidence to make some aspects of the proposed DeNS training strategy more clear for readers. I would like to increase my scores if the authors could address my concerns in this section and questions in the next section.
+
+### Questions
+1. In each iteration, the model uses either the standard training or the DeNS training. As DeNS training requires the forces to be encoded into the input atom features, I wonder how the force encoding module would be used in the standard training which instead uses the forces as labels.
+
+2. In Table 5, the authors demonstrate that the DeNS training is more efficient and results in larger performance gains than increasing max degrees of irreps. I wonder why the authors changed the model from EquiformerV2 to EquiformerV1 for this investigation. After all, the EquiformerV2 model is claimed to largely benefit from scaling the max degrees of irreps. How would EquiformerV2 with different max degrees of irreps behave in the same setting of Table 5?
+
+3. Could you provide more discussion on why you chose Equation 7 for the multi-scale noise scheduler?
+
+### Soundness
+2 fair
+
+### Presentation
+2 fair
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 3
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper proposes an auxiliary task (not a pre-training) to help learn molecular tasks: denoising not only equilibrium (few are available) structures, but also denoising the not-yet-at-mechanical-equilibrium structures (much more numerous).
+
+The key new idea is to provide also the forces (of the non-corrupted input) together with the corrupted structures, to make the problem well posed.
+
+Rather heavy experiments show that the proposed method can sometimes beat the state of the art.
+
+### Strengths
+Originality:
+
+The idea of the auxiliary task applied to non-equilibrium structures is great, here the paper allows to make this task well-defined by feeding the (uncorrupted) forces as input.
+
+
+Clarity:
+
+The paper explains carefully how the auxiliary task is added to the "default" model on which one wants to work.
+I like a lot figure 1, which explains very clearly and concisely the idea.
+
+
+
+Significance:
+
+I think the paper makes the point for releasing more non-equilibrium structures (although, this is already the case in OC20 and OC22 if I understood well), which is important to state clearly to the community.  As authors say:
+
+> We hope that the ability to leverage more from non-equilibrium structures as proposed in this work can encourage researchers to release data containing intermediate non-equilibrium structures in addition to final equilibrium ones.
+
+### Weaknesses
+Originality:
+
+The original idea is not groundbreaking: auxiliary tasks are known, the specific case of denoising as well, here the novelty is only in feeding also the forces as input.
+
+Quality:
+
+it is not always clear from the results shown in tables, that the proposed method improves the SOTA significantly.
+Also, since there are two measures of success (energy and forces), it's sometimes difficult to make a final decision.
+
+Clarity:
+
+The paper has some typos, but most importantly, experiments are discussed very quickly (too quickly). More space should be devoted to discuss results. For instance table 4 shows a nice improvement for DeNS on OOD splits (i.e. better generalization if I understand well that OOD is Out of Distribution as opposed to ID=In Distribution).
+
+Significance:
+
+Since the results are not strikingly better when using DeNS on the SOTA, and given it involves a number of additional hyperparameters (that obviously do not need very narrow fine-tuning, admitedly, but still, they involve more work and potential for problems), it is not clear yet whether the contribution would be used widely.
+
+Table 1d is probably the most convincing result, to me (differences between models seem more significant). Is it computed on the validation set(s) ? (or train set ? I have a doubt).
+
+In summary, this is red AI and the results are not significantly better than SOTA, thus not convincing for publication.
+
+### Questions
+I note that:
+- trainings are very heavy.
+- there is some hyper-param tuning, and DeNS is used on top of the best models (equiformerV2).
+- energy and force: sometimes only one is better, but as you say, it's a balance.
+- All that considered, most of the differences reported are not very significant. Maybe you could outline in green when a metric is significantly better in one model than in others (when this happens).
+Can you answer and/or improve the discussions (and/or presentation of results) in the Experiments section, to show that indeed, the improvement is significant ?
+
+If some tables show no significant improvement, it should be explained why.
+
+If the method is mostly able to speed up training to achieve equivalent accuracy, state it (and it will weaken the paper's claim, but at the same time strengthen the submission).
+
+
+
+Note:
+Table 2 does not show a significant improvement from using DeNS
+
+
+
+
+typos:
+devication -> deviation
+"L2 different" -> "L2 difference / L2 norm / squared difference "
+
+### Soundness
+2 fair
+
+### Presentation
+2 fair
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 4
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This paper proposes a method called denoising non-equilibrium structures (DeNS), which generalizes to a larger set of non-equilibrium structures without relying on additional datasets for pre-training, and the effectiveness of DeNS is demonstrated on the OC20, OC22, and MD17 datasets, achieving better results and faster training times compared to existing methods.
+
+### Strengths
+1. The inverse denoising framework presented in this work is quite interesting, as it proposes a novel possibility of combining structures, forces, and energy (or other properties) in a dual or inverse setting.
+2. The experiments and ablation study conducted are robust and extensive, accompanied by meticulous analysis.
+3. Well-written and easy to understand.
+
+### Weaknesses
+1. The results show limited improvements in the OC20, OC22, and MD17 datasets.
+2. Maybe could add more results from the denoising framework with other backbones to provide a comprehensive understanding of its performance.
+3. I believe that developing a general AI-based molecular dynamics (MD) method is more crucial than specifically designing and tuning for the OC dataset. A more generalized approach could be beneficial to the community by focusing on the broader application of deep learning-based MD methods across different systems, such as drug molecules, crsytal materials or polymers.
+
+### Questions
+1. If the improvements are not solely due to data augmentation, but rather are attributed to the inverse or dual denoising setting, it would be valuable to explore the generalization ability of the S2EF system in other systems.
+2. The tunable standard deviation (σ) denoising strategy appears to be crucial, it might be beneficial to incorporate visualizations of σ during the training process for better illustration.
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 5
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This paper proposes a new auxiliary task to improve the training of networks on the prediction of the energy and forces from the structure of an atomistic system. It proposes to consider the recovery of the 3d structure after perturbing it with Gaussian noise (similar to a denoising autoencoder). Yet, since this plain (structure-to-structure) denoising problem is only well-posed for equilibrium structures, the authors propose to consider the forces of the original structure as an additional input along with a corrupted non-equilibrium structure, which seems to make the denoising problem well-posed again. The resulting ability to also consider  non-equilibrium structures significantly increases the available amount of training data and results in significant and systematic numerical improvements on several data sets.
+
+### Strengths
+I am not familiar with machine learning on atomistic systems, such that my evaluation has to be treated with some care. Yet, 
+- The idea seems to be novel, very well motivated by language processing and computer vision, and nicely resolves the ambiguity of denoising non-equilibrium structures with the help of additional input. 
+- The numerical results (on huge and apparently very competitive datasets) are very promising as they are largely able to improve the state-of-the-art. 
+- Ablation studies over hyper-parameters, architectural choices, and loss functions indicate a very well-designed method.
+
+### Weaknesses
+The presentation of the work could be improved significantly. While part of my difficulty in understanding the presentation is surely due to my lack of knowledge in the particular field, I believe some aspects hold for scientific texts in general 
+- Abbreviations should not be used before they are introduced. The abstract already refers to "S2EF" and "IS2RE results" (clarified in the numerical results), refers to network architectures without citation and even uses variables ($L_\max = 2$ and $L_\max=3$) whose meaning is assumed to be known.
+- It is not explained why forces make the denoising problem on non-equilibrium structures well-posed and also not how the forces are obtained. Is it correct to assume that the potential of any structure can be computed and that the forces are the gradient thereof? If so, it would be just two additional sentences of explanation that make the work much more clear. 
+- The property "equivariant" is frequently used. While I know what it means, it was unclear to me what kind of equivariance (with respect to which transformations) is meant and why that influences the ability to encode forces. Rotations and translations of 3d coordinates/vectors? 
+- I do not know the term "type-L vectors" - what does it mean? I tried googling but did not have direct success. Thus, I'd recommend defining it. Also, what is "the projection of $f$ into type-L vectors with spherical harmonics"? Representing the function f with L-many coefficients corresponding to spherical harmonics basis functions?
+- It would have helped me to cite something for "SO(3) linear layer". 
+- Please double-check the manuscript for typos. 
+
+Please be aware that I am not part of your main audience. Thus, feel free to adjust the writing for those aspects where you believe your main audience would also agree.
+
+From my (quite uncertain) point of view, the strengths seem to clearly outweigh the suboptimal presentation (particularly if the latter is well suited for a more expert audience).
+
+------  
+As an update after the discussion phase, the points I thought were strong (contribution+numerical evaluation) are rather seen as critical by reviewers who are much more familiar with the field. As my initial rating was based on the assumption of a significant novelty, such that I am lowering my score to account for this.
+
+### Questions
+In addition to some small things raised above, I have two further questions:
+- It was strange to me to decide for a cost function for each batch during training with some probability. Will this not be equal to a weighted linear combination between the two terms in expectation?
+- The ablation in Table 1 (b) indicates that $p_{DeNS}=0.5$ is the best, but also the largest tested value. Wouldn't it make sense to ablate values $>0.5$?
+
+### Soundness
+3 good
+
+### Presentation
+2 fair
+
+### Contribution
+4 excellent

human_reviews/YFDM6uMMSE.md

@@ -0,0 +1,167 @@
+# EMOE: Expansive Matching of Experts for Robust Uncertainty Based Rejection
+
+- Decision: Reject
+- Scores: 5, 3, 5, 3
+
+## Abstract
+Expansive Matching of Experts (EMOE) is a novel method that utilizes support-expanding, extrapolatory pseudo-labeling to improve prediction and uncertainty based rejection on out-of-distribution (OOD) points. 
+We propose an expansive data augmentation technique that generates OOD instances in a latent space, and an empirical trial based approach to filter out augmented expansive points for pseudo-labeling. EMOE utilizes a diverse set of multiple base experts as pseudo-labelers on the  augmented data to improve OOD performance through a shared MLP with multiple heads (one per expert). We demonstrate that EMOE achieves superior performance compared to state-of-the-art methods on tabular data.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+3
+
+### Summary
+This paper proposes an expansive data augmentation approach, EMOE, to enhance uncertainty-based rejection on out-of-distribution data. EMOE comprises multi-stage training, i.e. initial expert learning and secondary matching model learning.  The novelty lies in the expansive data augmentation and filtering.
+
+### Strengths
+Expansive data augmentation and filtering are good points for this paper. Experiments have shown some good results.
+
+### Weaknesses
+This paper is not well written. The presentation of figures and methods is not clear. Fig.1 does not deliver any useful information about the approach. It should show multi-stage training and emphasize the difference between experts and heads. Fig.2 is also not informative, you can consider visualizing the selection on Alg.2. For eq.2-9, please clarify the meaning of g(x) nearby. Currently, g(x) is only shown in 165-167, which makes readers hard to understand these equations. Also, please include math symbols in Alg.1 as your paper defines.
+
+The experiment part is poor. Even the number of experts and the impact of $L_{mean}$ is not verified in experiments.
+
+Please give a clear definition of your problem at the beginning part of the method section. Besides, please clarify what is the meaning of distribution shift here. For example, distribution shifts can be covariate shifts in domain generalization or semantic shifts in out-of-distribution detection.
+
+### Questions
+Please refer to the weaknesses.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 2
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+3
+
+### Summary
+The authors propose a method for label generation by creating OOD instances in a latent space and performing empirical filtering. Specifically, EMOE is a pseudo-label generator equipped with multiple experts, and it demonstrates superior results over baselines on both CHEMICAL DATASETS and TableShift datasets.
+
+### Strengths
+1. EMOE introduces a novel support-expanding method in latent space to generate OOD instances, improving pseudo-labeling without additional OOD data.
+2. The method is data-modality agnostic, working across diverse data types (chemical, tabular, image) without domain-specific adjustments.
+
+### Weaknesses
+1. The authors mention extrapolatory capability in the introduction, which aligns with the goal of OOD generalization, but their proposed model does not directly address this issue.
+2. The normal distribution used in Equation 1 is overly strong and impractical; to validate this assumption, the authors should compare the OOD data generated by their method with real OOD datasets to assess the gap.
+3. The authors should conduct experiments on OOD generalization datasets, such as the full spectrum test of OpenOOD, to evaluate their model’s performance against other mainstream OOD methods.
+
+### Questions
+See Weaknesses
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 3
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+3
+
+### Summary
+This paper contributes to enhancing the ability of machine learning models to make predictions and reject uncertain questions when dealing with out-of-distribution data points. To enable model exploration, the authors propose a pseudo-label-based expansive matching method with multiple experts. Experimental results on multiple single-domain generalization benchmarks (i.e. Tableshift, chemical data and etc.) demonstrate the effectiveness of the proposed method.
+
+### Strengths
++ The problem of generalizing to OOD data is critical in many real-world applications, especially in medical and drug discovery domains. The paper's focus on improving extrapolation capabilities is timely and could have substantial practical implications.
++ Experimental results are conducted on multiple scenarios (e.g. tabular, chemical and etc), which demonstrate the generalization of the proposed method.
+
+### Weaknesses
++ Dependence on Base Experts: While the paper shows improvements over the base experts, there is a reliance on the quality and diversity of these base experts. The performance of EMOE could be limited if the base experts are not sufficiently robust or diverse. Meanwhile, the requirement of multiple experts also increases the computation complexity of the proposed method.
+
++ The experiments are conducted on simple MLP-based architecture. It would be better that the authors follow [1] and conduct experiments on the more powerful network architecture. For example, the empirical results in [1] show that most domain generalization methods fail to outperform the vanilla baseline when hyperparameters are carefully tuned based on the ResNet-50 and ResNet-101 architectures. It would be beneficial to adopt the network architecture and experimental settings from [1] for the PACS dataset to better demonstrate the effectiveness of the proposed method.
+
+[1] In Search of Lost Domain Generalization
+
+### Questions
++ Could the authors provide the qualitative results of the success and failure case during rejection? For example, the authors could visualize the image that the prediction are inaccurate and successfully or fail to reject. This could help the readers better understand the effectiveness and the limitations of the proposed method.
+
++ What is the value of 'k,' the number of experts used in the experiments? Is the rejection ability sensitive to the number of experts?
+
++ The paper primarily focuses on single-domain generalization scenarios. However, it is more common for training data to come from multiple domains. Another question is whether the proposed method can effectively generalize to multi-domain scenarios. Take PACS dataset as an example, the authors can take three domains from photo, art painting cartoon and sketch as the training data and the left one as the test set.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 4
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+The paper presents an approach to identify out-of-distribution (OOD) samples for tabular and sample data. It is proposed to form OOD samples on the feature space with augmentations and then pseudo-label them using a model ensemble. The proposed ensemble is a shared MLP backbone with several heads. In addition, there is a heuristic approach to filter out pseudo-labels that may not be useful.  In this case, a simple parametric model is used to evaluate the pseudo-labels. The approach is extensively evaluated on chemical and tableshifts datasets, as well as on image-based data from the PACS dataset.
+
+### Strengths
+- The proposed idea is simple and easy to understand. In general, the proposed approach is clearly described.
+- There is an extensive evaluation on chemical datasets. The method has been built for this type of data, so it shows that it works well in practice.
+
+### Weaknesses
+- The major issue of the paper is its limited novelty. It combines existing approaches in a straightforward manner. It does not show new elements even on the existing approaches. This point makes the work weak. The paper would stand easier to an applied ML venue. 
+- The whole approach and idea is better suited to domain generalisation. It would certainly make sense to present developments on domain generalisation.  The domain evaluation would be more appropriate. 
+- The PACS dataset is a domain generalisation dataset. It's not clear that it's being used in a different way. It would make more sense to choose a dataset for OOD detection if this is the goal of the paper.
+- Another major limitation is the luck of comparison with recent methods. It does not help the paper to show its potential. 
+- In Section 3.3, the L1 loss for classification is not well motivated. It should be argued why it makes sense to use it for the classification task, e.g. strong regularisation based on the literature. 
+- Line 178, the reference to "general data" adds ambiguity to the understanding of the approach. It would make more sense to write the paper on tabular data (excluding the images) and explain what the problems are there.
+
+### Questions
+- It would be interesting to explain why the paper was not written only for tabular data and focused only on domain generalisation. 
+- What is the reason to rely on PACS for OOD detection and pick up OOD datasets?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+1

human_reviews/aRxLDcxFcL.md

@@ -0,0 +1,203 @@
+# Virtual Community: A Generative Social World for Embodied AI
+
+- Decision: Reject
+- Scores: 5, 3, 5, 5
+
+## Abstract
+We present Virtual Community, a social world simulation platform designed to support embodied AI research, featuring large-scale community scenarios derived from the real world. Virtual Community introduces two key features to enrich the virtual social world with generative AI: scalable 3D Scene creation, which supports the generation of expansive outdoor and indoor environments at any location and scale, addressing the lack of a large-scale, interactive, open-world scene for embodied AI research; and embodied agents with grounded characters and social relationship networks, the first to simulate socially connected agents at a community level, that also have scene-grounded characters. We design two novel challenges to showcase that Virtual Community provides testbeds to evaluate the social reasoning and planning capabilities of embodied agents in open-world scenarios: Route Planning and Election Campaign. The Route Planning task examines the agent's ability to reason about time, location, and tools in the community to plan fast and economical commutes in daily life. The Election Campaign task evaluates an agent's ability to explore and connect with other agents as a new member of the community.  We evaluate several baseline agents on these challenges and demonstrate the performance gap of current methods in addressing embodied social challenges within open-world scenarios, which our simulator is designed to unlock. We plan to open-source this simulation and hope Virtual Community can accelerate the development in this direction.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4
+
+### Summary
+This paper introduces a novel method for creating 3D scenes at scale. They show how to use this method to generate scenes in Route Planning and Election Campaign.
+
+### Strengths
+The idea of generating new simulation environments with Generative AI seems very useful as it would accelerate the pace at which we create and test out RL agents in more complex environments than possible today. Virtual Community is an big innovative step towards this.  The paper is quite clear in its presentation (pending some typos) that make it easy to understand the contributions and impact it could have on the community. If done well and deployed well, Virtual Community could be significant in driving innovations in the world of RL by forcing researchers to think about novel algorithms that are required to become experts in such environments.
+
+### Weaknesses
+Results section feels lacking. Only results for Route Planning are included; it would be nice to see results for the Election Campaign as well. 
+But more importantly, the election campaign environment doesn't seem very useful. Route planning requires having an environment and a simulation since the agent is actively taking actions in that environment. However, the election campaign environment is more about interactions with other people - not something that immediately requires having a 3D environment.
+
+### Questions
+1. Only MCTS and LLMs are included for baseline performance. A2C today is standard enough to where it should be an algorithm that is benchmarked to further show how having this environment would be helpful to the community.
+
+2. How is the 3D environment helpful for the Election Campaign environment? Why is Virtual Community critical to generating this environment and what is an interaction that Virtual Community is enabling for this environment that wouldn't be possible with a simpler simulation of this? 
+
+3. If existing algorithms do so well on the current environments, what do you think are some unique insights that we could draw by using this environment and training RL agents on top of them? Is there an example of a "hard" environment that we could generate using Virtual Community.
+
+4. I don't fully understand what the complete end to end process for building an environment with Virtual Community looks like. Would be great if this were better explained in the paper.
+
+
+I am willing to change my rating if the above things are resolved!
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+---
+
+## Human Reviewer 2
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+5
+
+### Summary
+This work addresses the use of AI (including LLM) to endow virtual agents with intelligent behavior, which includes locomotion, interaction and some form of transactions. The authors appear to emphasize the specific constraints of embodiment as well as the inclusion of social relationship between agents. The paper puts forward two integrated tasks to illustrate its technical proposal: one, path planning, of intermediate abstraction level and the other, election campaigning, with a more complex relation to agents personality and social interactions.
+
+### Strengths
+Incorporating AI techniques into virtual agents platforms and simulation has been a longstanding goal, faced with numerous fundamental as well as practical challenges. This paper is thus ambitious in its exploration of the problem and timely in the method chosen, which investigates the potential impact of LLM on this endeavor.
+
+### Weaknesses
+This type of work has seen a resurgence in interest since the development of LLM [Wan et al., 2024], although it has a long history featuring symbolic AI techniques of search-based path planning and plan-based decision making and behavior, of which not much has found its way to the paper’s reference list of analysis of previous work. Although more contemporary techniques certainly show promise including for this sort of application, this is unfortunate as i) LLM are generally weak on relevant topics such as planning (despite recents experiments such as [Nasir et al., 2024]) and ii) some VR integration problems have been previously described and it does not seem appropriate to re-invent them: in particular the relationship between AI-based action parameters and the lower-level control of action and animation in the virtual world. 
+
+Path planning is one of the two main tasks allocated to these virtual agents. However, the corresponding sections suffer from both a lack of clarity and a lack of technical details. In particular there is no clear articulation between the low-level A*-based path planning and higher level (MCTS, LLM). This is particularly confusing since there is a potential overlap between various levels, including at the level of AI techniques (see, e.g. [Xiao et al., 2023]). Moreover, since the authors only mention A*, they are not addressing the well-known issue of dynamic environments, which becomes increasingly relevant with the number of agents and autonomous vehicles. 
+
+Following Figure 2, there is a lack of details on how social relationships are established from the characters’ profiles, whose description is relatively simplistic. Reference to “open-world knowledge” does not appear sufficient in the light of the vats body of work dedicated to persona definition with LLM. 
+
+Results of Table II are difficult to interpret, in particular in terms of LLM performance. In addition, 19 personal schedules and 106 commutes can hardly be considered a large-scale simulation in view of previous work. Some explication given for the results, such as leveraging environmental resources, is simply not useful in the absence of a fair comparison of integration of high-level decision making and low level decision across the three techniques (Rules, MCTS, LLM).
+
+Minor aside: There are some oddities in references, such as suggesting that Auerbach et al. [2014] would be a “foundational model”, in the sense of pretrained Transformers-based foundational models such as those listed in a previous section of the paper (Bubeck et al., 2023; Liu et al., 2023; Driess et al., 2023; Blattmann et al., 2023)
+
+
+References
+
+Wan, H., Zhang, J., Suria, A.A., Yao, B., Wang, D., Coady, Y. and Prpa, M., 2024, May. Building LLM-based AI Agents in Social Virtual Reality. In Extended Abstracts of the CHI Conference on Human Factors in Computing Systems (pp. 1-7).
+
+Aghzal, Mohamed, Erion Plaku, and Ziyu Yao. "Can large language models be good path planners? a benchmark and investigation on spatial-temporal reasoning." arXiv preprint arXiv:2310.03249 (2023).
+
+Song, C.H., Wu, J., Washington, C., Sadler, B.M., Chao, W.L. and Su, Y., 2023. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 2998-3009).
+
+Xiao, H. and Wang, P., 2023. Llm a*: Human in the loop large language models enabled a* search for robotics. arXiv preprint arXiv:2312.01797.
+
+Nasir, M.U., James, S. and Togelius, J., 2024. GameTraversalBenchmark: Evaluating Planning Abilities Of Large Language Models Through Traversing 2D Game Maps. arXiv preprint arXiv:2410.07765.
+
+### Questions
+How are social relations stablished from the agents' profiles. Are the actual agents' profiles more sophisticated than the examples given (e.g. through auto-generation or a memory for interactions, or evolving preferences?)
+How are the various decision-making methods (Rules, MCTS, LLM) integrated with lower-level functions like path planning?
+How are MCTS used in practice?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+1
+
+---
+
+## Human Reviewer 3
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4
+
+### Summary
+The paper proposes Virtual Community, a 3D social open-world simulator to evaluate current embodied intelligence research. The authors build a pipeline to automatically generate 3D interactive city scenes, which transform 3D geospatial data into simulation-ready scenes, and enrich the simulation platform with embodied characters in SMPL-like format with character informations and social communities. To showcase the significance of this simulator, the authors propose two tasks: Route Planning and Election Campaign, to benchmark the involving progress of current embodied agent.
+
+### Strengths
+1. The automatic 3D interactive city scenes generation pipeline is quite interesting and insightful, which consists of mesh simplification, texture refinement, object placement, and automatic annotation.
+2. It considers the rich personalities and social relationships that make up a community network, which may be meaningful for computational social science research.
+3. The presentation and visualization of the paper is quite good.
+
+### Weaknesses
+1. While the authors claim that the pipeline can generate an infinite number of scenes as shown in Table 1, it is still important to note how many assets this work has generated so far. It lacks a quantitative analysis of these created assets.
+2. The paper performs almost no quantitative experiments, only the proposed route planning task. This actually shows that the benchmark cannot cover too many current research methods, which is the biggest weakness of the paper.
+
+### Questions
+Since the paper uses human characters rather than actual humanoid or quadruped robots, what is the point of designing control spaces, physical properties, and interactive objects for them? It can be a navigation-only simulator.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 4
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4
+
+### Summary
+The paper presents Virtual Community, a generative social simulation platform designed for embodied AI research. It has two key contributions:
+1. Scalable 3D Scene Creation: This feature enables the automatic generation of expansive indoor and outdoor environments using real-world geospatial data, filling a gap in open-world simulation for embodied AI.
+2. Embodied Agents with Grounded Characters and Social Networks: The platform models agents with rich, grounded personalities and social networks, allowing for complex social interactions in large-scale, open-world settings.
+
+The paper also introduces two novel tasks—Route Planning and Election Campaign—to assess the social reasoning and planning capabilities of embodied agents within the platform.
+
+### Strengths
+1. The paper is well-organized and uses visuals effectively to clarify the platform's components and features.
+2. The platform models agents with grounded personalities and social networks, enabling the study of complex social interactions in a simulated community.
+3. The automatic generation of detailed 3D environments from real-world geospatial data, combined with generative models, is really impressive.
+
+### Weaknesses
+1. While the paper claims support for large-scale environments, it lacks quantitative results on how the system scales in terms of computation, memory, or performance metrics (e.g., agent count or scene complexity). There is also no mention of whether the simulator offers a ‘headless mode’ to allow multiple instances to run simultaneously on a cluster, which would be valuable for generating large-scale data or evaluating models in parallel. Additionally, it is unclear whether the simulator can be run ‘faster than real-time’ to speed up experiments. 
+
+2. The paper doesn’t explain how the movements of characters not being controlled by the baseline algorithms are managed. It is unclear whether these movements are based on a social forces model, predefined behavior heuristics, or something else. 
+
+3. The generative approach to creating agent personalities and routines may not fully capture complex or realistic social behaviors.
+
+4. The reliance on external geospatial data and APIs like Google Maps could limit applicability in areas with incomplete or unavailable data. Exploring methods for generating environments in a more data-independent manner would improve flexibility.
+
+5. The paper doesn’t discuss the usability of the simulator or the available APIs for developers. It’s also unclear whether there’s a built-in way to model interactions as graphs, or if users would need to design these graphs themselves. Providing tools to model and track interactions as graphs would make it more user-friendly. Clear documentation and API examples would also help developers integrate the platform into their projects more easily.
+
+6. There is no clear mention of the intended target audience for the simulator. 
+
+7. The paper omits references and comparisons to similar platforms like SEAN 2.0 [1] and Arena 1.0-4.0 [2], specifically along the axis of including the presence of robots in their simulator. 
+
+8. The platform focuses on simulation, but there is no discussion about how these simulated behaviors or learned policies can transfer to real-world applications. 
+
+[1] Tsoi, Nathan, et al. "Sean 2.0: Formalizing and generating social situations for robot navigation." IEEE Robotics and Automation Letters 7.4 (2022): 11047-11054.
+
+[2] Shcherbyna, Volodymyr, et al. "Arena 4.0: A Comprehensive ROS2 Development and Benchmarking Platform for Human-centric Navigation Using Generative-Model-based Environment Generation." arXiv preprint arXiv:2409.12471 (2024).
+
+### Questions
+Refer to the Weaknesses section.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3

human_reviews/aZ1gNJu8wO.md

@@ -0,0 +1,139 @@
+# A Geometric Framework for Understanding Memorization in Generative Models
+
+- Decision: Accept
+- Scores: 8, 8, 6
+
+## Abstract
+As deep generative models have progressed, recent work has shown them to be capable of memorizing and reproducing training datapoints when deployed. These findings call into question the usability of generative models, especially in light of the legal and privacy risks brought about by memorization. To better understand this phenomenon, we propose the \emph{manifold memorization hypothesis} (MMH), a geometric framework which leverages the manifold hypothesis into a clear language in which to reason about memorization. We propose to analyze memorization in terms of the relationship between the dimensionalities of $(i)$ the ground truth data manifold and $(ii)$ the manifold learned by the model. This framework provides a formal standard for ``how memorized'' a datapoint is and systematically categorizes memorized data into two types: memorization driven by overfitting and memorization driven by the underlying data distribution. By analyzing prior work in the context of the MMH, we explain and unify assorted observations in the literature. We empirically validate the MMH using synthetic data and  image datasets up to the scale of Stable Diffusion, developing new tools for detecting and preventing generation of memorized samples in the process.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+3
+
+### Summary
+Memorization in Generative model occurs when model reproduce its training data instead generalising from it. This phenomenon is undesirable because it represent to modelling failure. The paper introduces the "manifold memorization hypothesis" (MMH) to explain memorization geometrically. It suggests that memorization happens when the manifold learned by the model at a point has a lower dimensionality than the true data manifold at that point. This framework defines two types of memorization:
+Overfitting-Driven Memorization (OD-Mem): This occurs when the model assigns low dimensionality to a region of the manifold, overfitting specific training data points.
+Data-Driven Memorization (DD-Mem): This happens when the true data distribution inherently has low dimensionality in certain regions, leading the model to reproduce specific details in these regions even if it generalizes well.
+
+The paper introduces the Manifold Memorization Hypothesis (MMH) as a geometric framework to analyze and understand memorization in deep generative models (DGMs). It explores memorization through the lens of local intrinsic dimension (LID), a metric used to assess whether a model's manifold dimension aligns with that of the underlying data. The authors identify two types of memorization: overfitting-driven memorization (OD-Mem) and data-driven memorization (DD-Mem), categorizing instances of memorization based on LID discrepancies. This hypothesis is validated empirically with synthetic and real datasets, demonstrating how MMH can predict memorization tendencies and proposing methods to mitigate memorization in diffusion models, particularly Stable Diffusion.
+
+### Strengths
+The introduction of MMH offers a fresh perspective on memorization, connecting it to geometric properties and adding depth to existing theoretical understandings.
+ The paper proposes actionable techniques to reduce memorization in generative models, which has direct implications for improving model safety and compliance.
+
+### Weaknesses
+The reliance on geometric and measure theory could make it challenging for a broader audience to fully understand and appreciate the framework. 
+There should be more experiments
+While LID proves to be an effective measure, there is some overlap between memorized and non-memorized samples in LID estimates, which may affect its reliability in distinguishing memorized samples in real-world applications.
+
+### Questions
+How does the MMH framework extend to discrete generative models, such as those in NLP (transformer)?
+Could alternative LID estimation methods reduce the overlap issue observed in the experiments, or is this an inherent limitation of the current approach?
+Are there specific types of datasets or model architectures where MMH may not apply effectively?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+4
+
+---
+
+## Human Reviewer 2
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+3
+
+### Summary
+Memorization is a big issue in deep generative models. Recent works have highlighted that such models are prone to duplicating their training points, given certain prompts, or generating samples that appear very close to training points with slight modifications (e.g., background and color). To combat and understand this phenomenon, this paper proposes a geometric framework which leverages the manifold hypothesis. Particularly, the work analyzes memorization in terms of the relationship between dimensionalities of the ground truth data manifold and the manifold that is learned by the model. This is done by analyzing and calculating the Local Instrinsic Dimension (LID) of the data and generated points, and building a theoretical understanding around the model's memorization behaviors. Overall, the paper validates the applicability of the manifold analysis in the detection of memorized samples through the development of new tools and prevention methods in generating such samples.
+
+### Strengths
+The paper provides a very valuable and interesting perspective about memorization, which comes in two forms: Overfitting-driven Memorization (OD-Mem) and Data-driven Memorization (DD-Mem). These two forms of memorization highlight that memorization is a convoluted aspect of trained deep networks, and it should not always be naively thought of as just overfitting. The paper does a decent job with making a case for this topic, by using LID as a metric of identifying memorized samples. 
+
+Furthermore, the work introduces a differential metric, $\mathcal{A}^\text{FLIPD}$, which allows for the optimization of a conditional variable $c$ to prevent the reproduction of training points during sampling. Specifically, the authors extended and combined the work of [Wen at al. (2024)](https://arxiv.org/pdf/2407.21720) and [Kamkari et al. (2024)](https://arxiv.org/pdf/2406.03537) to produce the aforementioned metrics, which show promising results in the prevention of memorization for stable diffusion models.
+
+### Weaknesses
+The comparison of various LID metrics to other memorization detection metrics is not quite clear, especially in figure 6. It is difficult to tell how much more effective LID metrics are since the authors did not contrast them well in their written text, and it is difficult to interpret the results in the density histograms of fig. 6 without showcasing what is classified as memorization, visually.
+
+Furthermore, it is still unclear on how to quantitatively differentiate between OD-Mem and DD-Mem samples. The description, visualization (fig. 1), and toy example (fig. 4) certainly highlight the differences, but for natural datasets, it is difficult to tell still.
+
+### Questions
+For definition 3.1, could you elaborate on condition (2) or why $p_\theta$ cannot be equal to $p_*$?
+
+Could you elaborate the differences between $s_\theta$ and $s_\theta^\text{CFG}$ in equation 41 of the Appendix D.2? 
+
+What is the conclusion of figure 6? The importance of this figure/result is not clear to me. 
+
+Could you provide more figures like figure (a), of memorized samples, in Figure 5? What was the threshold used in the detection metric from [Carlini et al. (2023)](https://arxiv.org/pdf/2301.13188), which was used to identify memorized samples?
+
+Both $\hat{LID_\theta}$ and $\hat{LID_*}$ are estimators of LID. Is it fair to assume that $\hat{\text{LID}}_*$ is more accurate? Or, is that the wrong interpretation?
+
+### Soundness
+4
+
+### Presentation
+2
+
+### Contribution
+3
+
+---
+
+## Human Reviewer 3
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+3
+
+### Summary
+The paper proposes MMH, a geometric framework for understanding memorization in generative models by focusing on the dimensions of learned manifolds. It aims to analyze memorization as a function of the ground truth data manifold and the learned model manifold. The framework distinguishes between two types of memorization: overfitting-driven and data-driven, providing a theoretical base and empirical tools for detecting memorization across a range of models, from synthetic data to image generation using Stable Diffusion.
+
+### Strengths
+1. The authors combine the manifold memorization hypothesis and a geometric perspective to memorization, potentially unifying existing observations into a cohesive framework.
+2. The paper provides clear motivation and background on memorization in DGMs, connecting memorization to legal and ethical implications.
+3. The empirical experiments well support the propositions and theorems.
+
+### Weaknesses
+1. Some minor typos, for example ''subtantial'' should be ''substantial'' in line 44.
+2. The MMH concept does not introduce substantial novel contributions beyond covering DD-Mem rather than only considering OD-Mem. This work is more like summarizing the existing work with their framework. For example, Proposition 3.5 is a (in)-formal summary of the findings in [1]'s experiments with MMH.
+
+[1] Yuxin Wen, Yuchen Liu, Chen Chen, and Lingjuan Lyu. Detecting, explaining, and mitigating memorization in diffusion models. In Proc. ICLR, 2024.
+
+### Questions
+1. I am confused about the definition of LID. From the whole paper, I can tell the smaller the LID is the more memorization. In line 134, it is defined as ''the number of degrees of freedom'' instead of the number of variables needed in a minimal representation of the data. However, based on lines 125-126, it seems to be the same as the intrinsic dimension. Could you clarify?
+2. In line 182: ''a notable consequence is that it cannot be detected by comparing training and test likelihoods''. Can you elaborate more about this?
+3. As mentioned in the weaknesses, could you explicitly state your novel contributions more, or provide specific examples of how the framework goes beyond existing work? Currently, there is no comparison with the existing frameworks, such as [2]. I will be happy to increase my score if this is resolved.
+
+[2] Robi Bhattacharjee, Sanjoy Dasgupta, and Kamalika Chaudhuri. Data-copying in generative models: a formal framework. In International Conference on Machine Learning, 2023.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+3

human_reviews/c43FGk8Pcg.md

@@ -0,0 +1,189 @@
+# Denoising Diffusion Step-aware Models
+
+- Decision: Accept
+- Scores: 6, 6, 6, 8
+
+## Abstract
+Denoising Diffusion Probabilistic Models (DDPMs) have garnered popularity for data generation across various domains. However, a significant bottleneck is the necessity for whole-network computation during every step of the generative process, leading to high computational overheads. This paper presents a novel framework, Denoising Diffusion Step-aware Models (DDSM), to address this challenge. Unlike conventional approaches, DDSM employs a spectrum of neural networks whose sizes are adapted according to the importance of each generative step, as determined through evolutionary search. This step-wise network variation effectively circumvents redundant computational efforts, particularly in less critical steps, thereby enhancing the efficiency of the diffusion model. 
+    Furthermore, the step-aware design can be seamlessly integrated with other efficiency-geared diffusion models such as DDIMs and latent diffusion, thus broadening the scope of computational savings. Empirical evaluations demonstrate that DDSM achieves computational savings of 49\% for CIFAR-10, 61\% for CelebA-HQ, 59\% for LSUN-bedroom, 71\% for AFHQ, and 76\% for ImageNet, all without compromising the generation quality.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This paper proposes Denoising Diffusion Step-aware Models (DDSM), which utilize variable-sized neural networks for different steps of the diffusion generative process. The key insight is that diffusion steps have varying importance, so uniformly allocating computational resources is inefficient. The method trains a slimmable UNet that can be flexibly pruned to different capacities. An evolutionary search then determines the optimal per-step network size to balance efficiency and performance. Experiments demonstrate substantial computational savings on CIFAR-10, CelebA-HQ, LSUN-bedroom, AFHQ, and ImageNet versus conventional diffusion models, with minimal quality loss.
+
+### Strengths
+- The core ideas are technically sound and offer a unique perspective on accelerating diffusion models.
+- Using evolutionary search to determine step-wise network requirements. 
+- Compatibility with existing methods like DDIM.
+
+### Weaknesses
+- Using a different network at different timestep has been explored before, such as in e-diff-i. 
+- The compatibility claims with DDIM and latent diffusion are fairly cursory. More detailed experiments showing accelerated performance combining these methods could better showcase modularity.
+- The evolutionary search itself requires non-trivial compute resources. Analysis of the search costs and scalability could be insightful.
+
+### Questions
+You claim compatibility with methods like DDIM and latent diffusion, but details are limited. Could you provide in-depth quantitative experiments demonstrating accelerated performance when combining DDSM with these existing diffusion acceleration techniques?
+
+The search cost and scalability of the evolutionary algorithm is unclear. Could you analyze the computational requirements of the search procedure and discuss how it scales with factors like dataset size?
+
+### Soundness
+3 good
+
+### Presentation
+2 fair
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 2
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper...
+- proposes to accelerate diffusion sampling by using diffusion models of different size at each time-step,
+- proposes to use evolutionary search to find a best step-aware strategy,
+- shows the effectiveness of the proposed approach on CIFAR-10, CelebA-HQ, and ImageNet sampling.
+
+### Strengths
+- Paper is easy to follow.
+- Using networks of different size (slimmable network) for each-time step is an unexplored approach in diffusion acceleration.
+- DDSM shows promising performance gains on a variety of generation tasks.
+
+### Weaknesses
+While the idea behind DDSM is interesting, I am inclined to give "marginal reject" due to weak experimental validation.
+
+- The paper lacks comparison with [1], which I think is a very relevant work.
+- The paper lacks experiments on higher resolution data. Can the authors provide results on $\geq 128$ resolution images?
+
+[1] Structural Pruning for Diffusion Models, Fang et al., NeurIPS, 2023.
+
+### Questions
+- Figure 3 is unclear. What do networks with mixed colors (blue and green) mean?
+- In Table 3, why do DDSM outperform ADM-large? Shouldn't the performance of DDSM be bounded by the performance of the largest model?
+- Is DDSM compatible with guided diffusion? Can the authors provide some demonstrations?
+- Is DDSM compatible with recent fast solvers, such as EDM [2]?
+
+[2] Elucidating the Design Space of Diffusion-Based Generative Models, Karras et al., NeurIPS, 2022.
+
+### Soundness
+2 fair
+
+### Presentation
+2 fair
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 3
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.
+
+### Summary
+The paper proposes Denoising Diffusion Step-aware Models (DDSM) to improve efficiency of denoising diffusion probabilistic models (DDPMs) for image generation. Previously, DDPMs require compute-intensive iterative sampling, using the full model each step. DDSM hypothesizes different steps have varying importance, and uses a spectrum of networks with adapted sizes for each step, determined via evolutionary search. This avoids redundant computation in less critical steps.
+DDSM integrates with slimmable networks - trained simultaneously on sub-networks to enable execution at arbitrary sizes.
+
+### Strengths
+DDSM accelerates diffusion models by avoiding uniform computation for all steps. The step-aware network design is shown to be efficient and effective across datasets.
+
+### Weaknesses
+The major concern is the pathway of denoising acceleration.  It's evident that the prevalent approach to enhancing diffusion speed, in the context of network compression, hinges on the application of post-training quantization techniques. These techniques enable the compression of neural networks in a manner that circumvents additional training [1,2]. However, I observe that your method necessitates training and, notably, falls short in performance when compared to methods that forego training. To illustrate, empirical assessments show that DDSM facilitates computational reductions of 49% for CIFAR-10 and 76% for ImageNet. In contrast, the techniques in [1,2] manage to achieve 4 or 8-bit quantization (surpassing DDSM in speed) without compromising the FID score. It's crucial to underscore that these methods achieve this efficiency entirely without the need for further training. Consequently, the approach employed by DDSM for accelerating denoising doesn't appear to be robust enough.
+
+
+[1] PTQ4DM, CVPR 2023
+
+[2] Q-dfiifusion, ICCV 2023
+
+### Questions
+Refer to weaknesses.
+
+### Soundness
+3 good
+
+### Presentation
+2 fair
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 4
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper presents a novel approach to accelerating sampling from denoising diffusion models without sacrificing image quality. This approach is based on the idea that not every generation step is equally difficult nor equally important for the quality of the final image. The authors propose using *slimmable networks* to prune the network weights at less-important generation steps, and combine this with an evolutionary search method to identify which steps can be safely pruned without sacrificing image quality. The authors demonstrate that their method reduces the FLOPs required for multiple datasets while achieving comparable or better sample quality (FID) than the unpruned methods, and further show that it can be combined with the adaptive step-skipping of DDIM or with latent diffusion for further improvements. They also visualize the learned pruning schedules for different datasets, and show that the set of safely-prunable generation steps differs depending on dataset structure.
+
+### Strengths
+**[S1]** The idea of pruning the network adaptively based on the difficult of each generation step is insightful, and well-motivated by the introduction and the pilot study in section 3.1. This approach also seems quite novel. I'm not aware of any previous work that has considered adaptively using different network sizes at different timesteps.
+
+**[S2]** The authors show that their approach is actually complementary to existing acceleration methods like DDIM, and yields additional improvements when combined. I expect that the approach could also be combined with some of the more recent ODE-based acceleration methods like DPM-Solver, since the proposed DDSM technique is agnostic to the exact mathematical sampling process and instead adaptively shrinks the network itself.
+
+**[S3]** The empirical results are quite strong across multiple datasets. The authors show significant cost savings relative to sampling from the full model, and actually show slight *improvements* in sample quality for all but one of the datasets (LSUN-bedroom takes a small hit).
+
+**[S4]** The approach is easy to understand and seems like it would not be too difficult to implement. The authors also plan to release their code.
+
+**[S5]** I found the results of the pruning schedule search to be quite insightful. Different datasets show different amounts of pruning, and this pruning occurs at different times, which seems to match whether the dataset has more information in its high-frequency or low-frequency components.
+
+### Weaknesses
+**[W1]** The approach relies on "slimmable networks" and "slimmable counterparts" of standard convolution and normalization. However, these aren't discussed in much detail. I think the paper could benefit from some additional background on what slimmable networks are and how they work, and more details on the particular slimmable network architecture used for this work. (In particular, how many slimmable switches were used, and how do they fit into the U-Net architecture?)
+
+**[W2]** The objective for the evolutionary search algorithm is also presented at a fairly high level and could use more details. The authors mention using NSGA-II to "balance conflicting objectives", but it's not clear to me what objectives were used. Algorithm 2 suggests that the search objective was a linear combination of FID and FLOPS, but the details of the linear combination weights are not specified.
+
+**[W3]** Although the contributions of this work do seem orthogonal to some of the more recent work in accelerating diffusion model sampling, I think the experimental results would be more impressive if they could also be demonstrated for these more recent accelerated sampling approaches like DPM-Solver, [DPM-Solver++](https://arxiv.org/abs/2211.01095) or the recent [UniPC](https://arxiv.org/abs/2302.04867). My guess is that the methods could be combined, but it would be useful to see how much FLOPs can be saved when combined with these more recent samplers.
+
+### Questions
+Could you provide more details on the slimmable network architecture and on the configuration of the search algorithm, as I discuss in [W1] and [W2]? For the search algorithm, what was the value of $w_M$ used, and how critical is this choice? (Or, are the FID and FLOPs measurements automatically balanced by NGSA-II somehow?)
+
+Figure 4a and Figure A are quite interesting, but they also look somewhat blurry. I can sort of see some fuzzy "bands" at different points of the trajectory; are these actual changes in the slimmed model size or are these some sort of compression artifact in the figure image? Also, what is the resolution of the X axis here, e.g. where are the boundaries between different steps? It's hard to understand exactly what the plot is showing, and I think this plot would be more readable if it were presented as a line graph (instead of as a heatmap), since its only 1 dimensional.
+
+I found some very recent related work ["Structural Pruning for Diffusion Models" (Fang et al. 2023)](https://arxiv.org/abs/2305.10924) which was just accepted at NeurIPS. That work also considers pruning in order to speed up diffusion model inference, although I believe they only consider pruning the entire network rather than adaptively pruning at different timesteps. It might make sense to discuss this related work in your paper, and I'd also be curious how your approach compares to theirs in terms of FLOPs savings.
+
+Have you explored combining your DDSM approach with some of the more recent accelerated sampling approaches? And do you still observe runtime improvements for those methods?
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good

human_reviews/cKkZEj8n7O.md

@@ -0,0 +1,191 @@
+# Generalization Error Minimized Deep Learning
+
+- Decision: Reject
+- Scores: 5, 3, 5, 3
+
+## Abstract
+Despite the vast applications and rapid development of deep learning (DL), understanding and improving the generalization ability of deep neural networks (DNNs) remains a fundamental challenge. To tackle this challenge, in this paper, we first establish a novel bias-variance decomposition framework to analyze the generalization error of DNNs. Based on our new generalization error formula, we then present a new form of DL dubbed generalization error minimized (GEM) DL by jointly minimizing the conventional optimization target and an analytical proxy for the generalization error. Extensive experimental results show that in comparison with DNNs trained within the standard DL, GEM DNNs have smaller generalization errors and better generalization ability, thereby improving DNN prediction accuracy. Notably, GEM DL can increase prediction accuracy by as much as 13.19% on ImageNet in the presence of data distribution shift between training and testing.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+3
+
+### Summary
+This paper introduces a new algorithm of DNN training named generalization error minimized (GEM). It is motivated by theoretical analyses of reducing the generalization error. Targeting on minimizing a squared generalization error, the authors derive that the squared expected test loss and expected squared test loss are the two key components for reducing generalization error. GEM is designed based on this theoretical result. Experiments CIFAR-100 and ImageNet confirm the effectiveness of GEM against ERM and DOM. Additional experiments of test data shift, few-shot learning and imbalanced classification also show GEM surpasses ERM.
+
+### Strengths
+1. This paper proposes a new algorithm of DNN training derived from generalization error minimization. 
+2. The algorithm is easy to implement and doesn’t require heavy computations. 
+3. GEM shows clear benefits in small datasets.
+
+### Weaknesses
+1. The theoretical part is a bit tricky to me. Eventually the generalization error is reduced to be two loss terms only relying on test data as shown in Eq (15). Then the algorithm is implemented by simply replacing the test data with training data or processed training data. For the latter case, the difference between training and test data must be known in advance, which is not very realistic. Also, the theoretical result is based on the particular form of squared generalization error, which is different from the common definition of generalization error (training loss - test loss). The motivation of using the squared form of generalization error should be explained.
+2. The optimal choices of beta and lambda are quite different for different datasets. This increases the complexity in applying the proposed loss in real applications. In the experiments, how did you decide the optimal choices?
+3. For the evaluation of Case-2, how is ERM implemented? I think it’s not fair to compare GEM with a vanilla ERM trained on clean data if the test data shift is supposed to be known. Simply data augmentation (e.g., image compression or Gaussian Blur) should work better. 
+4. For few-shot learning settings, transfer learning is a common technique to address the issue of data limitation. Training from scratch is far from state-of-the-art. It’s interesting to see whether GEM also improves the accuracy when a pre-trained model is utilized for few-shot learning.
+
+### Questions
+See Weaknesses
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 2
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+This paper introduces a framework that integrates a 'novel' bias-variance decomposition of generalization error into the training objective for deep neural networks (DNNs). The proposed framework aims to improve the model's generalization by jointly optimizing the traditional ERM loss and a theoretically-inspired proxy for generalization error, yielding gains in accuracy  across various image classification scenarios.
+
+### Strengths
+1- The paper is extremely well-written and structured. 
+2-  Generalization in deep learning is a very important problem. The proposed generalization decomposition is simple, yet sound and interesting.
+
+### Weaknesses
+W1- From a theoretical perspective, the generalization decomposition is quite straight-forward for me. So the main contribution here lies in leveraging the theoretical bound to propose a practical proxy of generalization. Regarding this proposed proxy, I have some concerns (See the question bellow) regarding the link between the theory and practice.
+
+W2- The experiments focused only on image classification tasks with CNN-based structure with NLL loss. I would like to have seen a more diversified  experiments, e.g.,  transformer-based models. regression tasks, NLP, to better evaluate the utility of the proposed approach. 
+
+I will reconsider my score if the authors can address my concerns in the questions section.
+
+### Questions
+**Theory**
+
+In eq 11, the first two terms use the test data (True population). The third term is based on the training data and the forth uses both distribution. In the remaining part of the paper (after eq 11 Section 4.1), the authors argue that we can neglect the last two terms and keep only the first 2 (defined using the population data T ). However, in the experiment section, the proposed generalization proxy is based on the training data D and not on independent data). This poses several contradictions: 
+
+Q-1: The proposed proxy does not minimize  $\text{Var}(\Omega(T, \hat{\theta}) | \hat{\theta}) + K^2(\hat{\theta})$. In fact, as it is based on the training data, what it really minimizes is $\text{Var}(\Omega(D, ) | \hat{\theta}) + KJ^2(\hat{\theta})$, which contradicts the argument made by the author (between eq11 and 13) that these two terms should be ignored. 
+
+Q-2:  With the proposed proxy setup, T and $\hat{\theta}$ are no longer independent. In fact $\hat{\theta}$ is optimized using T. Hence, Eq 4 and onwards are no longer valid. 
+
+Q-3: One potential way to solve this in my opinion is to define the proxy using an auxiliary validation set, i.e., in equation 19, the first term is defined using the training data while the second and third are computed using an independent validation data. 
+
+
+**Experiments**
+
+Q-4: Even that the link between the theory and the generalization proxy is not valid or weak (Q1-3 above), I think that the proposed regularizer can still be useful. However. the empirical validation in this paper is not enough to evaluate that. Have the authors tried their regularizer with other architectures beside CNN and NLL loss, e.g., transformer?
+
+Q-5. Empirically, the authors compare only with DOM. However, several other theoretically-sound approaches have been proposed in the literature that can reduce overfitting, e.g., dropout [1-2], Mixup [3-4], feature diversity [5-7], etc. 
+
+
+[1] Srivastava, Nitish, et al. "Dropout: a simple way to prevent neural networks from overfitting." The journal of machine learning research 15.1 (2014): 1929-1958. \
+[2] Wan, Li, et al. "Regularization of neural networks using dropconnect." International conference on machine learning. PMLR, 2013.  \
+[3] Zhang, Hongyi. "mixup: Beyond empirical risk minimization." arXiv preprint arXiv:1710.09412 (2017). \
+[4] Zhang, Linjun, et al. "How does mixup help with robustness and generalization?."  Neurips (2020). \
+[5] Laakom, Firas, et al. "WLD-reg: A data-dependent within-layer diversity regularizer." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 37. No. 7. 2023. \
+[6] Laakom, Firas, et al. "Learning distinct features helps, provably." ECML, 2023. 
+[7] Cogswell, Michael, et al. "Reducing overfitting in deep networks by decorrelating representations." ICLR 2016
+
+### Soundness
+1
+
+### Presentation
+3
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 3
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+2
+
+### Summary
+This paper introduces generalization error minimized (GEM) deep learning, a method that optimizes the standard minimization objective coupled with a “generalization” proxy target. The authors first decompose the conditional generalization error into three terms: Conditional testing variance, Conditional training variance and bias between the expected training and testing performance. The authors then discuss two cases: Case 1 when the training and test distribution are similar and Case 2 where the train and test distribution are dissimilar. In case 1, the GEM objective boils down to the linear combination of cross entropy (CE) loss, square of the CE loss and the second moment of the CE loss. Results with GEM in image classification task on CIFAR, ImageNet with ResNet, MobileNetV2 and other backbones shows that GEM performs better than ERM.
+
+### Strengths
+The authors show experimental results on a wide range of settings include few-shot setup, imbalanced learning in the context of image classification. The results on various datasets and settings show that GEM performs better than ERM in most of the scenarios.
+
+### Weaknesses
+1. I am not convinced the by the Case 2 argument. In most cases, the distribution (U, V) can not be obtained by processing (X, Y). For example, consider the case of subgroup shift (Waterbirds, Celeb-A) or OOD datasets (ImageNet-Sketch, ImageNet-R etc). 
+2. The authors should also report performance on various OOD datasets atleast in the case of ImageNet (for e.g ImageNet-A, ImageNet-Sketch to name a few)
+3. It would be great if the authors can also show a toy experiment with a synthetic data and how the solution learned by GEM and ERM differs
+
+### Questions
+1. Is label smoothing used in both ERM and GEM? If yes, what is the alpha value? The pseudocode in Appendix A.2 mentions label smoothing. 
+2. In case of Case 2, the authors should compare against a method that computes CE loss w.r.t $\hat{X}$ and original CE loss and not the general ERM.
+3. Could the authors provide the detailed hyperparameter and regularization configurations in each of the setups. In L524-529, the authors mention that many regularization techniques are already included in the baseline method. I could not find the relevant details anywhere in the paper.
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+---
+
+## Human Reviewer 4
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+This paper proposes a new algorithm called generalization error minimized deep learning (GEM DL). The algorithm builds upon a bias-variance decomposition for the generalization error. The decomposition then inspires a new form of regularization term to be added to the objective function. The authors empirically show models trained using the new algorithm experience better generalization performance compared to the naive empirical risk minimization.
+
+### Strengths
+- The paper studies a very interesting topic to the research community.
+- Overall, the type of experimentations that are done to evaluate the effectiveness of the proposed objective function are good. There are however issues within the experiments which I’ll detail in the weaknesses section.
+
+### Weaknesses
+- Overall, the paper is difficult to read. There are various vague expressions/statements throughout the paper. Starting from the abstract, what is meant by “a new form of deep learning”? Or in the introduction “training/testing performance”? This ambiguity makes it difficult to follow the authors' approach, especially when specific details are introduced before fundamental concepts are defined. Further contributing to the confusion, Section 3.1 refers to "average loss" as "performance", which contradicts the conventional understanding of these terms.
+- The motivation for introducing new definitions of generalization error and conditional generalization error remains unclear. The authors should clarify why they deviate from established definitions used in theoretical and empirical research. A detailed comparison between their proposed definitions (Equations 2 and 3) and conventional ones is necessary, highlighting the differences and justifications for these changes. Also, the definition in Equation 2 is more the “generalization gap” rather than the “generalization error”.
+- Ultimately, the proposed optimization method reduces to adding a regularization term to the training loss (or empirical risk minimization objective function).  Therefore, comparisons with benchmark regularization algorithms are crucial. However, the empirical results are limited in this regard, with only one benchmark included. Moreover, this chosen competitor exhibits lower performance than baseline ERM on the ImageNet dataset (Table 2), suggesting it may not be a suitable benchmark for state-of-the-art comparisons.  Furthermore, most of the experiments omit this competitor entirely, presenting only comparisons to ERM, which is insufficient. Even if considered a "plug-and-play" method, the proposed algorithm should be compared against other techniques that incorporate regularization terms into the objective function.
+- The application of the proposed method in Case 2 relies on a questionable assumption: knowledge of the distribution shift that will occur in the test case. In practice, both the type and amount of shift are often unknown a priori, raising concerns about the method's practicality in out-of-distribution scenarios. This limitation is evident in Figure 1, where the applied GEMs are pre-designed for the specific type of distribution shift. Moreover, when the amount of shift is unknown and an incorrect hyperparameter is estimated (e.g., overestimating the blurring amount with sigma=3 in Figure 1.b), performance deteriorates significantly compared to baseline ERM. This experiment highlights the method's sensitivity to accurate estimations of the distribution shift.
+
+
+More minor weaknesses/comments:
+- The statement "since  ˆθ depends only on D and the training process, it follows that T and ˆθ are independent" is inaccurate. Model parameters depend not only on the training data itself but also on the underlying training data distribution, which may not be independent of the test data distribution. The authors need to clarify the distinction between the independence of random sampling used to collect training and test data, and the potential dependence between the underlying distributions from which these samples are drawn.
+- The proposed method is adding two more hyper-parameters to the learning method, which would require additional hyper-parameter selection, making the method less practically appealing. 
+- There are no standard deviation bars in any of the figures throughout the paper. This is particularly problematic in Figure 4 where there are lots of small variations.
+
+### Questions
+- In Section 3.1, the defined generalization error appears to be its out-of-distribution variant. The authors should clarify the rationale behind using this specific setting and explain why they refer to this generalization gap as a "new generalization error." What novelty does it introduce? Additionally, the presence of a squared term within the generalization error definition requires justification.
+
+### Soundness
+1
+
+### Presentation
+2
+
+### Contribution
+1

human_reviews/cphhnHjCvC.md

@@ -0,0 +1,216 @@
+# End-to-End (Instance)-Image Goal Navigation through Correspondence as an Emergent Phenomenon
+
+- Decision: Accept
+- Scores: 8, 6, 6, 8
+
+## Abstract
+\vspace{-0.1cm}
+
+Most recent work in goal oriented visual navigation resorts to large-scale machine learning in simulated environments. The main challenge lies in learning compact representations generalizable to unseen environments and in learning high-capacity perception modules capable of reasoning on high-dimensional input. 
+The latter is particularly difficult when the goal is not given as a category (``\textit{ObjectNav}'') but as an exemplar image (``\textit{ImageNav}''), as the perception module needs to learn a comparison strategy requiring to solve an underlying visual correspondence problem. This has been shown to be difficult from reward alone or with standard auxiliary tasks.
+We address this problem through a sequence of two pretext tasks, which serve as a prior for what we argue is one of the main bottleneck in perception, extremely wide-baseline relative pose estimation and visibility prediction in complex scenes. The first pretext task, cross-view completion is a proxy for the underlying visual correspondence problem, while the second task addresses goal detection and finding directly. We propose a new dual encoder with a large-capacity binocular ViT  model and show that correspondence solutions naturally emerge from the training signals. Experiments show significant improvements and SOTA performance on the two benchmarks, \textit{ImageNav} and the \textit{Instance-ImageNav} variant, where camera intrinsics and height differ between observation and goal.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The authors introduced pretext tasks and a dual visual encoder for ImageNav and Instance-ImageNav navigation in 3D environments, which provide rich geometric information and make it possible to address the challenging mono-view setting with end-to-end trained methods. Via experiments it is shown that peformance on competing methods and SOTA on both benchmarks is better.  The idea of cross-view completion and goal direction computation as pre-text for ImageGoal in contrast to ObjectGoal is the philosophical novelty in the work. Apart from that, the pipeline and architecture presented will aid future scope of research in the benchmark challenges.
+
+### Strengths
+Fig. 3 is well described to give the view of the problem scope being breaked into cross view completion, relative pose estimation and visual navigation.
+This is a well written paper with clear explainations and technical soundess.
+The results sections and ablation studies uphold the claims.
+This paper will help the ImageGoal community - hence recommended.
+
+### Weaknesses
+The last portion of supplementary video in terms of correspondence needs better representation and also that is the core focus.
+The analyis of time complexity for doing correspondances in the benchmark and the distribution of work load should have helped understand the bottlenecks for a near real time system like robotic agents, even in embodied setups.
+A practical deployment in robotic setup should have confirmed the real world transfer applicability.
+I think Fig. 1 image you ahve search to related the chair with big picture - can any other image or better reoslution be used?
+Same with panaromic and mono view - please help in making sense of the image if at all included in body. If space contsraint, appendix referral is there for later sections, but introduction setup has to be clear.
+I think related work can only focus on the core related work, getting rid of object goal and general visual nav in such detail - this space can be used elsewhere in explaination later.
+No limitations of the work is presented. Future gaps should be explained well.
+
+### Questions
+"we split this path into 5 parts cor" - any logic regarding the discreet steps evenly spaced?
+Instead of Active Neural SLAM, anything else has been tried out in the pipeline?
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 2
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.
+
+### Summary
+This paper is concerned with image-goal navigation and proposes a new end-to-end method for this task. The main contribution of the work is the inclusion of a pre-training stage of two auxiliary tasks that help in learning relevant visual features to be used as a state representation to a recurrent policy. The method achieves state-of-the-art performance on publicy available datasets.
+
+### Strengths
+The paper is well written and easy to follow. The authors show that they understand the underlying challenges of this problem well and clearly explain their thought process behind the proposed approach.
+
+I agree with the authors that the bottleneck for these navigation problems is perception. I think the pretext tasks being proposed help in learning relevant representations for this task which would be otherwise very difficult to learn in typical end-to-end methods. Overall I think the paper proposes  a novel and sensible approach and moves the needle forward on this subject and improves the sample efficiency of these methods.
+
+### Weaknesses
+The title is somewhat misleading as the emergence of correspondences is really not the focus of the work, but more of an afterthought. The content of the paper might be mistaken as an investigation into this phenomenon that is carried out in this paper: 
+[A] Tang et al, Emergent of Correspondences from Image Diffusion, arXiv, 2023.
+In fact the emergence of correspondences from the learned representation is not really that surprising, given that the pre-training task is relative pose estimation. Even before the introduction of transformers, monocular pose estimation methods have shown that the representation learns to identify meaningful keypoints on objects. One example:
+[B] Mousavian et al, 3D Bounding box estimation using deep learning and geometry, CVPR 2017
+
+I appreciate the inclusion of the experiment where DEBiT was integrated with ANS for a direct comparison to a modular approach. However, I am surprised by the large performance gap to the proposed method. It seems unintuitive that a method that tries to map pixels directly to actions would outperform an approach that uses a map and de-couples the planning from the control. I think this should be looked more carefully to ensure fair comparison. Was the global policy of ANS finetuned with the frozen binocular encoder b and using the AdaptFormers? ANS was trained for object-goal so probably the Neural SLAM and global policy components need to be re-trained for the ImageGoal task. As it stands, the experiment is not convincing that end-to-end methods are better than modular approaches.
+
+### Questions
+Why is using panoramas treated as unrealistic with regards to robotic applications? RGB sensors (and even RGB-D) are relatively cheap and can be easily mounted on robots to cover a 360 degree field-of-view.
+
+The authors cover some of the literature on pre-text tasks for improving sample efficienty but I think these are also worth discussing:
+[C] Ye et al, Auxiliary Tasks and Exploration Enable ObjectGoal Navigation, ICCV 2021
+[D] Sax et al, Mid-Level Visual Representations Improve Generalization and Sample Efficiency for Learning Visuomotor Policies, CoRL 2019
+Especially regarding [D], shouldn't a representation trained on a multitude of vision tasks be more task-generalizable to just pre-training on relative pose estimation?
+
+### Soundness
+3 good
+
+### Presentation
+4 excellent
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 3
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper improves the view correspondence encoding in image-goal navigation by pre-training a large binocular ViT model (DEBiT) with cross-view completion (CroCo) and relative pose estimation and visibility prediction (RPEV) from images in photorealistic indoor scenes. Specifically, the dual encoder in DEBiT takes binocular images as input, and the encoded features are merged with a decoder to model the correspondence between the two images. DEBiT is pre-trained with CroCo, followed by RPEV to estimate the relative distance, relative rotation, and overlap between two images. When addressing downstream Instance-ImageNav and ImageNav tasks, DEBiT takes the goal image and agent's observation as inputs, its network is adapted in tuning, and the output representations will be passed to the policy network for decision-making. Results show substantial improvement compared to previous approaches.
+
+### Strengths
+- This paper studies an important problem in visual navigation: learning high-capacity perception modules for modeling and reasoning view correspondence. It proposes DEBiT, which implements a dual encoder to enable early fusion of the images. It introduces CroCo and RPEV pre-training, which are highly relevant and have shown to be effective in addressing the problem. 
+- Significant improvement is achieved compared to previous approaches, boosting the ImageNav and Instance-ImageNav results to 94% SR and 59.3% SR, respectively. The methods introduced in this paper and the resulting pre-trained models are very likely to inspire/to be used by future research in relevant fields.
+- The paper is technically sound; many important arguments and design choices are justified by experiments. Besides the key results, there are many highly constructive analyses/findings, such as "directly training binocular encoder from scratch", "tuning visual encoder with downstream policy network", "early fusion vs. late fusion", "attention with scale changes", etc., that are very valuable to future research.
+- Overall, this paper is nicely written; it is very compact, informative, and clear. All methods (and most of the implementation), visualizations, and discussions are clearly presented.
+
+### Weaknesses
+- The proposed DEBiT model seems to be limited in addressing image-goal navigation tasks (the image-nav task itself needs more justification), and it is unclear how it might benefit other visual navigation problems.
+    - However, I do believe that CroCo and RPEV can help in learning a general (and better) navigation-specific perception model (e.g., for obj-Nav, language-guided-Nav, Audio-Nav, etc.), while I am concerned that it might be much less effective compared to ImageNav which is defined to take two images as input.
+    
+- I am aware that ImageNav is an interesting visual navigation problem and has gained some research attention (publications), but I am still not convinced by its setting, especially how much practical value it might bring to real-world applications. [This question might be more appropriate for researchers who proposed ImageNav, but since this paper is devoted to ImageNav, I believe it must have a very strong reason and motivation behind.]
+    - The two ImageNav tasks consider indoor short-range navigation for finding static objects (Instance-ImageNav has paths with an avg. geodesic length of 12.41m); from the user's perspective (e.g., a household robot), giving instruction in the form of a specific image is unnatural.
+    - From the ImageNav data statistics and the visualization provided in this paper (supp. video), it seems that most of the paths have very few intersections, and the agent doesn't really need to explore the environment but keeps moving to new regions until the target is in its sight. This could be the reason why a very simple policy network, without methods like SLAM, can still lead to amazingly high performance (94% SR). I am concerned that the ImageNav task itself oversimplifies the problem, and this paper might overclaim the contribution to visual navigation research.
+
+- Some experiments are missing to justify arguments (see Questions below).
+
+- The limitation and future extension of this work are not discussed in this paper.
+
+### Questions
+Questions without a star (*) are not critical to my evaluation of this paper, but I still hope the authors can kindly and briefly respond to them. Please also respond to my concerns in Weaknesses.
+
+- (*) This paper mentioned depth inputs, but I wonder why depth images are not applied in the model since it might greatly facilitate learning view correspondence and identifying space and obstacles. (I might have overlooked some details; please correct me if I did.)
+
+- (*) How does the choice of $\tau$ influence the pre-training and the downstream results? If two images are too distant away and have little overlap, will it be too noisy to learn? Or it might help the visual encoder to learn distant image correspondence and benefit exploration in navigation? Any numerical analysis on this?
+
+- (*) The paper claims the benefit of having visibility estimation but does not quantify its impact on downstream tasks. Any results for this?
+
+- Just curious, can CroCo and RPEV be trained simultaneously? 
+
+- In Table 2, from Tiny to Large, the model size increases drastically, but the difference in navigation results is quite small; what might be the reason? Does it mean that a very large perception model might not be necessary? What would be the result of DEBiT-Tiny + Adapters?
+
+- Instance-ImageNav depicts targets viewed from a different camera; if this is the main reason why the proposed method gets much less improvement compared to ImageNav (an OOD situation as claimed), I wonder what if the images for pre-training are augmented with different camera parameters, (1) Is it still feasible to learn CroCo and RPEV? (2) Will it reduce the visual domain gap in downstream?
+
+- Are the visual encoders in OVRLs fine-tuned with the policy networks in downstream tasks? From their papers and Table 4, it seems Yes. But in the section Related Work, "Once pre-trained, the encoder is often frozen before passing into a policy learning module."
+
+- About comparison experiments.
+    - Using adapters in the perception model largely improves the results. I wonder how much improvement it might bring to the previous approaches.
+    - In Table 3, for a fair comparison (in terms of pre-training tasks and #params), I think DEBiT-Tiny and RPEV-only should be listed, which I believe is more rigorous and can prove the same argument.
+    - (*) In Table 4, an important point to mention is the size of the visual encoders (#params) applied in each work. After scanning some papers in the Table, I believe there is a clear trend of larger-models-better-results. But it seems that DEBiT is relatively efficient, especially compared to OVRL-v2. I hope the authors can clarify this point, which I believe will also strengthen the argument.
+
+Other Suggestions:
+- Remove "(Instance)-" from the title.
+- I found a paper, "Learning navigational visual representations with semantic map supervision (Hong et al., ICCV2023)", which also focuses on learning better navigation visual encoder with view correspondence and uses two images for pre-training. It seems relevant and could be added to the references.
+
+### Soundness
+4 excellent
+
+### Presentation
+4 excellent
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 4
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.
+
+### Summary
+In this work, the authors introduce pretext tasks and a dual visual encoder for ImageNav and Instance-ImageNav navigation, offering rich geometric information and enabling the resolution of the challenging mono-view scenario with end-to-end trained techniques. 
+
+The method breaks down the problem into multiple training stages, demonstrating the emergence of solutions to the correspondence problem without explicit supervision. 
+
+Through experiments, the manuscript highlights the effectiveness of the proposed pretext tasks and a dedicated dual encoder architecture, surpassing competing methods and achieving  a novel state-of-the-art performance on both benchmarks. Additionally, the authors demonstrate seamless integration into a modular navigation pipeline of the proposed approach.
+
+### Strengths
+- In the state of the art (Section 2), a detailed review of all the literature related to the problem investigated in this work is conducted. Furthermore, the main differences of the proposed model with respect to previous works are outlined.
+
+- The system proposed in Section 2 is novel. It describes an integration of subsystems/submodules that allows for an approximation to the problem, yielding highly promising results. Additionally, the proposed design can be considered innovative. The integration of Cross-View Completion (CroCo) into a navigation system as described has not been explored before, although it's true that this contribution is not the strongest of all. The second pretext task, relative pose estimation and visibility (RPEV) for navigation settings, is interesting and provides significant results. Overall, the originality of the work is considerable.
+
+- The experimental evaluation provides very interesting results in which the proposed model outperforms the state of the art (see tables 4 and 5). An ablation study is provided where the real impact of each of the contributions or parts of the model can be interpreted. Finally, qualitative evidence is presented on how the proposed pre-training process results in the emergence of correspondences between images by analyzing the attention of the last layers.
+
+### Weaknesses
+The experimental evaluation included in the paper has some minor limitations that should be addressed:
+- It is not clear on which databases or environments the system has been tested. For example, on what environments are the results reflected in Table 5 obtained? The interested reader has to review (Krantz et al., 2023) to know these details. Section 4, "Experimental setup" subsection needs to be improved.
+
+- I believe one of the most interesting contributions of the proposed model is that it generates correspondences between images due to how the pre-training is designed. Figure 4 shows some results or qualitative evidence in this regard. However, the paper does not explain in detail how this image is generated or how these correspondences are analyzed. The manuscript simply mentions that they "visualize averaged attention of the last cross-attention layer of a DEBiT-L model", but more details should be provided.
+
+### Questions
+Overall, I see an strong paper here.
+I would simply suggest to the authors that they provide explanations for the limitations I have described in the previous section.
+
+### Soundness
+4 excellent
+
+### Presentation
+3 good
+
+### Contribution
+3 good

human_reviews/djM3WzpOmK.md

@@ -0,0 +1,192 @@
+# Neural Snowflakes: Universal Latent Graph Inference via Trainable Latent Geometries
+
+- Decision: Accept
+- Scores: 8, 5, 8, 6
+
+## Abstract
+The inductive bias of a graph neural network (GNN) is largely encoded in its specified graph. Latent graph inference relies on latent geometric representations to dynamically rewire or infer a GNN's graph to maximize the GNN's predictive downstream performance, but it lacks solid theoretical foundations in terms of embedding-based representation guarantees. This paper addresses this issue by introducing a trainable deep learning architecture, coined \textit{neural snowflake}, that can adaptively implement fractal-like metrics on $\mathbb{R}^d$. We prove that any given finite weighted graph can be isometrically embedded by a standard MLP encoder, together with the metric implemented by the neural snowflake. Furthermore, when the latent graph can be represented in the feature space of a sufficiently regular kernel, we show that the combined neural snowflake and MLP encoder do not succumb to the curse of dimensionality by using only a low-degree polynomial number of parameters in the number of nodes. This implementation enables a low-dimensional isometric embedding of the latent graph. We conduct synthetic experiments to demonstrate the superior metric learning capabilities of neural snowflakes when compared to more familiar spaces like Euclidean space. 
+
+Additionally, we carry out latent graph inference experiments on graph benchmarks. Consistently, the neural snowflake model achieves predictive performance that either matches or surpasses that of the state-of-the-art latent graph inference models. Importantly, this performance improvement is achieved without requiring random search for optimal latent geometry. Instead, the neural snowflake model achieves this enhancement in a differentiable manner.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+The authors leverage the representation power of a snowflake metrics using learnable neural snowflake activation in a neural network-based model. In doing so, they prove that neural snowflake permits universal representation for learning latent graphs, which is not the case for simpler previous state-of-the-art geometric embeddings. This is demonstrated theoretically via Theorems 1-2, with particular examples provided to motivate a learnable, non-Riemannian metric. This is also demonstrated empirically with a comprehensive comparison to alternative latent graph inference methods which use more typical metrics, clearly outperforming these baselines in some cases, and performing competitively in others.
+
+### Strengths
+1. The benefit of the extension to quasi-metric spaces and a learnable neural snowflake activation is clearly explained, as the authors provide both theoretical findings and concrete examples to demonstrate the utility.
+2. The experiments are thorough and well-controlled. The striking improvement in applying neural snowflake in the synthetic setting is convincing, and the marginal benefit in the "input graph" experiment of Table 3 demonstrates that this advantage is present in real data as well.
+3. The universal representation power in conjunction with the fact that the model provably needs not to use a large number of parameters is a very strong theoretical result.
+
+Overall, this is a very well-written and convincing work. While I am not entirely knowledgeable on the topic of latent graph inference, I think that the theoretical and empirical arguments of the authors provide a very clear explanation of the novelty of their approach and how it fits in the context of the current state of the field.
+
+### Weaknesses
+While the majority of the experimental results are convincing, the results of Table 4 are slightly underwhelming. It is certainly a good finding that neural snowflake is competitive in all cases, which cannot be said about any of the other approaches, but the fact that it is usually outperformed by some other method limits the applicability of this approach to real data.
+
+Maybe there is some experiment to evaluate the method in a way that provides more initial information than the experiment in Table 4, but not entirely the original input graph as used in Table 3? However, I don't believe this is entirely necessary to convey the utility of the approach.
+
+### Questions
+While you note that this is mainly tangential to the primary focus of this paper, you mention that the suboptimality of the Gumbel Top-k edge sampling algorithm may be a reason why the improved performance of your method in the real data settings is not as pronounced. However, would this not affect all methods equally? That is, why do you suggest that an improvement in this regard might better differentiate the neural snowflake model from the baselines?
+
+### Soundness
+4 excellent
+
+### Presentation
+4 excellent
+
+### Contribution
+4 excellent
+
+---
+
+## Human Reviewer 2
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This work studies learning latent structures from pure point clouds (a set of points without graph structures) or graph-structured data (a set of points with observed structures). The authors present rogirous analysis into the property of latent structure learning models and how to conduct latent graph inference that satifies the metric properties. Experiments on several benchmark datasets verify the proposed approach.
+
+### Strengths
+1. The paper studies an important and interesting problem
+
+2. The theoretical analysis is rigorous and with technical depth
+
+3. The overall writing is good though some parts can be improved to increase the readability
+
+### Weaknesses
+1. One concern lies in the clarity of the proposed model. The paper title describes the model as a "universal" method, but the main paper does not justify this point very well. For example, how universal the model is and to what extent? What is the advantage of the proposed model compared with prior art? Also, the algorithm for the model is missing, which makes the reader hard to understand what is precisely done in the model implementation.
+
+2. The implication of the theoretical results needs to be made more clear. How can the theory apply to practical problems and what is the insight behind the analysis? 
+
+3. The experiments are limited with four small datasets. For the Cora and Citeseer, it seems that the common benchmark settings are not used for evaluation. Can the authors provide more reasons on this and what precisely the data splits are used for the present experiments? The baselines for comparison are not sufficient, and some of the typical graph structure learning models are missing, e.g., LDS [1] and the state-of-the-art NodeFormer [2] (that also considers the Gumbel trick).
+
+[1] Luca Franceschi et al., Learning discrete structures for graph neural networks. In ICML, 2019.
+
+[2] Qitian Wu et al., NodeFormer: A Scalable Graph Structure Learning Transformer for Node Classification. In NeurIPS, 2022.
+
+### Questions
+See some of the questions in the weakness section. Besides, there are some further questions:
+
+1. Can the model handle more difficult datasets, e.g., heterophilic graphs and graphs with incomplete edges?
+
+2. How does the model compare with recent graph structure learning models LDS [1] and NodeFormer [2]?
+
+3. What is the impact of the informativeness of node features on the model performance when not using the input graph?
+
+4. Can the theoretical results be applied to graph Transformers whose attention networks can be seen as latent graph inference?
+
+### Soundness
+3 good
+
+### Presentation
+2 fair
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 3
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+The paper presents and analyzes a trainable metric termed "neural snowflake", which is a combination of a bounded gaussian kernel distance , a fractal component $\vert a-b \vert^\alpha$ with $\alpha \in (0,1]$ and an *irregular* fractal component in log space $\log\left(1+\Vert a-b\Vert^\beta\right)^{1+\vert p \vert}$ with $\beta \in (0,1],p\in \mathbb{R}_0^+$. This distance function is used to learn *latent* graphs from data, wherein the graph structure of a dataset or sample is inferred from the features, instead of being constructed by a human annotater.
+
+The paper proves that
+
+1. parametrized with an MLP as embedder and the neural snowflake and distance function, we can learn a $d$ dimensional embedding for any finite weighted graph
+2. There are finite weighted graphs which are *not* similarly representable if using a euclidean distance
+3. the depth and hidden width of the MLP depends favourably on the number of nodes and embedding dimensions (order $(n\log n)^{1.5}$ and linearly, respectively)
+4. For specific graphs (e.g. trees) even more favourable guarantees are possible
+
+The method is evaluated on Cora, CiteSeer (standard GNN benchmarks) and Tadpole and Aerothermodynamics (used in the paper which introduced the latent graph approach.)
+
+### Strengths
+The paper is very accessible for such a technical topic and presents its story clearly and convincingly.
+
+
+Novelty/Originality: Moving from euclidean and Riemanian metrics to quasi-metric spaces is an idea which I haven't seen a lot in the literature ...and I checked because I was actually thinking about going into this direction myself but hadn't properly started working on this due to previous projects. Happy to be "scooped" in this manner!
+
+Significance: Having provable representation power in this sense for arbitrary graphs will be an important ingredient as we move towards building less supervised relational ML models
+
+Quality: derivations, expositions and experiments are well presented and clean. Proofs look sane to me as well, although I did not step through them in detail
+
+### Weaknesses
+1.  There is relatively high variances in most of the tables, could you please perform suitable significance tests (e.g. with a bonferonni correction) and mark those that are significant to aid the viewer?
+2. I might have missed it, but what is the STD across? Evaluations or trainings? E.g. in table 7
+3. Nitpick, but explicitly discussing the memory constraints and other concerns in a limitations section in the appendix helps the readers judge potential pitfalls
+
+### Questions
+Aside from Q2 in weakness, no questions
+
+### Soundness
+4 excellent
+
+### Presentation
+4 excellent
+
+### Contribution
+4 excellent
+
+---
+
+## Human Reviewer 4
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+The authors present Neural Snowflakes, a latent graph inference method. Neural Snowflakes are iterative representations that use a trainable activation function to refine input distance values $\|x - y\|$. The authors present theoretic justification for their model as well as compelling empirical results on latent graph inference tasks.
+
+### Strengths
+* The authors present compelling empirical results for their method
+* Theoretical analysis underpins these empirical results
+* The fact that Neural Snowflakes can break the curse of dimensionality and don't need exponentially many parameters is a very nice property
+
+### Weaknesses
+* The paper does not state the loss function the Neural Snowflake models are trained on
+* There is no mentioning of the computational complexity/ scalability of the approach: does it scale to hundreds of thousands/ millions of data points? Why (not)?
+
+### Questions
+Admittedly, I am not an expert on latent graph inference, so I'm doing my best to judge the quality of the work based on the description by the authors and the results produced.
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good

human_reviews/dyrGMhicMw.md

@@ -0,0 +1,210 @@
+# Initializing Models with Larger Ones
+
+- Decision: Accept
+- Scores: 8, 6, 8, 6
+
+## Abstract
+Weight initialization plays an important role in neural network training. Widely used initialization methods are proposed and evaluated for networks that are trained from scratch. However, the growing number of pretrained models now offers new opportunities for tackling this classical problem of weight initialization. In this work, we introduce weight selection, a method for initializing smaller models by selecting a subset of weights from a pretrained larger model. This enables the transfer of knowledge from pretrained weights to smaller models. Our experiments demonstrate that weight selection can significantly enhance the performance of small models and reduce their training time.  Notably, it can also be used together with knowledge distillation. Weight selection offers a new approach to leverage the power of pretrained models in resource-constrained settings, and we hope it can be a useful tool for training small models in the large-model era.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper introduces a novel approach to initialize smaller models given large pretrained models. The approach is simple, efficient and shows good results compared to random init and other numerous baselines and many image classification tasks.
+
+### Strengths
+1. The problem of obtaining strong weights for small models given large pretrained models is important and solving it efficiently is challenging.
+
+2. The paper explores a very interesting idea of simple weight copying from the source to target network as an efficient and training-free way to initialize smaller model. 
+
+3. The uniform selection approach is reasonable and takes into account multi-head structure and group convolution of source networks.
+
+4. The paper has many nicely outlined and run experiments that overall look interesting and convincing enough.
+
+5. The approach outperforms random init and works well on different vision tasks. 
+
+6. The paper is well presented and enjoyable to read.
+
+### Weaknesses
+1. The paper misses some important related works:
+- [Czyzewski2022] Breaking the Architecture Barrier: A Method for Efficient Knowledge Transfer Across Networks
+- [Chen2021] bert2BERT: Towards Reusable Pretrained Language Models
+- [Chen2015] Net2net: Accelerating learning via knowledge transfer
+
+Specifically, [Czyzewski2022] introduced a very similar weight transfer technique that seems more general than in the submitted paper since it allows for the source and target architectures to be different. They copied the center slices from source weights, which could be an interesting ablation in this submission. [Chen2015] proposed a simple approach to replicate weights to initialize models. While [Chen2015,Chen2021] focus on initializing a larger model, the idea of copying the weights from the source network to a target one is very similar and should be discussed in detail. This makes the novelty of the paper limited.
+
+2. The paper also should report results of student models pretrained and transferred to smaller datasets. While this would not be a fair baseline to the proposed method since it requires expensive pretraining, it could give important clues about what is the best performance of the model on this task. In addition, some papers (e.g. [Knyazev2021]) reported that very little pretraining is required to achieve competitive transfer learning performance. Therefore, it's not clear if the proposed weight selection approach would be competitive if someone has resources to pretrain a small target net on some large data first (at least for a little bit).
+
+- [Knyazev2021] Parameter Prediction for Unseen Deep Architectures
+
+3. For structured pruning, there are many stronger methods than L1 pruning, e.g. [Kwon2022, Halabi2022] and the baselines therein. The results in Table 6 are quite close and L1 pruning is also very simple, therefore stronger pruning methods could potentially match the performance.
+
+- [Kwon2022] A Fast Post-Training Pruning Framework for Transformers
+- [Halabi2022] Data-Efficient Structured Pruning via Submodular Optimization
+
+4. Most of the experiments are performed on transferring weights from ViT-S to ViT-T which have the same number of layers (12). It would be more interesting to see results when the source model is also deeper (which is usually the case in practice). Related to that, the teacher size results show that ViT-S is the best teacher and a potential reason for that could be that the layer selection approach (first-N) is not the best (perhaps a uniform selection of layers similarly how it is done for weights could be better than first N?). In Table 7, it would be very interesting to see the respective results for distillation which usually works better if the source model is larger/stronger (see the DeiT paper).
+
+Despite all the weaknesses, I'm looking forward to the the authors' response and will be willing to update the rating.
+
+### Questions
+1. In Net2net [Chen2015], for initialization the copied weights are rescaled to take into account the width difference (i.e. for wider/thinner networks one needs smaller/larger weight values according to standard weight initialization such as K.He or X.Glorot). Have the authors tried to rescale copied weight values?
+
+2. For "distill + weight selection" is the target model first initialized with weight selection and then the distillation is run or it's different?
+
+3. It seems that the results of baselines (random init and mimetic init) are different from previous papers like the mimetic init paper even though the setup looks the same. For example, in the mimetic init paper for CIFAR-10 ViT-T the accuracy is 87.39 for rand init and 91.38 for mimetic init, while in this submission it is 92.4 and 93.3 respectively. Can the authors elaborate on the reasons for that?
+
+4. In L1 pruning experiments, is the final number of selected weights the same as in the proposed weight selection? It's a little bit surprising to see that L1 underperforms uniform weight selection implying that the latter is actually a stronger pruning method. Can the authors elaborate on that?
+
+### Soundness
+2 fair
+
+### Presentation
+4 excellent
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 2
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This work introduces a method aimed at accelerating the convergence of training by transferring pre-trained knowledge from a different network. The approach involves extracting well-trained weights, which exhibit characteristics similar to high-pass filtering (HPF) or low-pass filtering (LPF), from a pretrained network and transplanting these weights onto a novel network, as opposed to starting with random initialization. This technique facilitates quicker training of the new architecture. The authors conduct comparisons between this method and random initialization, demonstrating its superior performance.
+
+
+However, I have some concerns as follows:
+
+1)	Firstly, it's important to note that the comparisons made in this work may not be entirely adequate. From my understanding, this approach falls somewhere between traditional initialization and transfer learning. Given that the method involves adopting a set of pre-trained filters or weights, it can be argued that this work is closer in nature to transfer learning rather than a simple initialization process. Therefore, it might be more appropriate to compare it with transfer learning methods rather than just random initialization. If the proposed method can approach or match the convergence levels of ImageNet pre-training, it would further support its practicality and utility in various applications, reinforcing its value in the field of deep learning.
+
+2)	Comparing the proposed method directly to random initialization may not provide a fair assessment, given the distinct differences in the initialization processes. Unlike the proposed method, a network initialized randomly must construct its structural filters (HPF or LPF) from scratch, which typically requires some epochs and computational resources. In contrast, the proposed method can bypass this initial construction step and is expected to be faster.
+
+3)	I strongly recommend reconsidering the categorization of this work from merely weight initialization to a more encompassing concept of knowledge transfer to a novel architecture. This work goes beyond traditional weight initialization methods by borrowing well-structured filters from a pre-trained network with a different architecture. This transfer of knowledge to a distinctly different architecture represents a significant departure from conventional weight initialization and indeed underscores the novelty of the approach. Categorizing it as a knowledge transfer method would better capture its unique contribution in the field.
+
+4)	Building upon the suggested change in categorization to knowledge transfer, it raises an interesting question about the potential applicability of the proposed method to highly dissimilar architectures. For instance, exploring the feasibility of transferring knowledge from a pre-trained ResNet to a Vision Transformer (ViT) is a compelling idea like Model Evolution. Training ViT models often necessitates access to large-scale datasets, but they can offer superior performance in various tasks. If the proposed method could indeed facilitate the transfer of knowledge from a ResNet to a ViT, it would hold significant promise. This potential application could lead to advancements in leveraging the strengths of different architectures and addressing challenges related to dataset size, thus expanding the scope and impact of the proposed method.
+
+I appreciate your insights, and it's clear that you find the work promising. However, I think that the manuscript doesn’t well categorize the proposed method and the experiments are not adequately conducted. I believe this work can be much improved.
+
+### Strengths
+See above
+
+### Weaknesses
+See above
+
+### Questions
+See above
+
+### Soundness
+3 good
+
+### Presentation
+2 fair
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 3
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper introduces 'weight selection', a simple but effective method that uses the weight of pre-trained large model to initialize the small model. The authors have conducted extensive experiments, showcasing the method's superiority compared to initializing models from scratch.
+
+### Strengths
+1. The experiments are comprehensive, covering both vision transformers and CNNs. Additionally, they have delved into various aspects such as knowledge distillation, comparison with pruning methods, and linear probing, which offer valuable insights
+2. The paper is easy to read and follow.
+
+### Weaknesses
+See questions.
+
+### Questions
+1. Could the authors give further explanation on the difference between 'consecutive selection' and 'uniform selection' in section 3.2?
+2. For comparison with pruning, could the authors elaborate on how L1 pruning was implemented? I am asking the question since the embedding dimension is changed.
+
+### Soundness
+3 good
+
+### Presentation
+4 excellent
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 4
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper suggests that initializing the weight parameters of a smaller neural network from a pre-trained, larger model can significantly improve the smaller model's performance.
+
+### Strengths
+- This paper is easy to follow;
+- The proposed method can improve the performance.
+
+### Weaknesses
+- The proposed method of selecting layers or weights for initialization is somewhat trivial. A more technical approach is encouraged to fully leverage the capabilities of larger models.
+
+- The study is limited in its scope by only employing ViT-T and ConvNeXt-F architectures. Inclusion of additional architectures such as Swin, PVT, and EfficientViT would enrich the evaluation.
+
+- There is an inconsistency in the reported accuracies for the same ImageNet-1k on the same model across different tables. For instance, ConvNeXt-F yields an accuracy of 76.1% in Table 1 but achieves around 82.0% in Table 2. Clarification is needed.
+
+- A comparative analysis involving ConvNeXt is crucial in the "COMPATIBILITY WITH KNOWLEDGE DISTILLATION" section, given its architectural differences with ViT.
+
+- The paper lacks sufficient motivation and theoretical grounding for the proposed selection strategy. More insights are needed to justify why, for example, the first N layers are selected, and how this methodology generalizes to other architectures.
+
+- For better comprehension, it would be beneficial to present the customized structures, such as ConvNeXt-F, in tabulated form.
+
+- The potential for leveraging the weights of larger models (which are more powerful) is high. However, Table 7 indicates that the proposed method struggles to effectively learn from these larger models.
+
+- Some grammar errors: 
+	- "Glorot & Bengio (2010a) maintains " -> "Glorot & Bengio (2010a) maintain";
+	- "Lin et al. (2020) transforms" -> "Lin et al. (2020) transform";
+	- etc.
+
+### Questions
+Please see Section "Weaknesses".
+
+### Soundness
+2 fair
+
+### Presentation
+2 fair
+
+### Contribution
+3 good

human_reviews/e9YuyOaJbc.md

@@ -0,0 +1,188 @@
+# Optimal and Generalizable Multimodal Representation Learning Framework through Adaptive Graph Construction
+
+- Decision: Reject
+- Scores: 3, 3, 3, 5
+
+## Abstract
+Multimodal contrastive learning train neural networks by levergaing ata from heterogenous sources such as images and text. Yet, current multimodal learning architectures cannot generalize to an arbitrary number of modalities, need to be hand-constructed and are often not robust to missing modalities. We propose AutoBIND, a novel contrastive learning framework that can learn representations from an arbitrary number of modalities. AutoBIND uses a graph-based approach to automatically select the most correlated modalities and uses a contrastive loss to learn the representations. AutoBIND is robust to missing modalities as it can dynamically update the graph structure during training. Based on the proposed framework, each modality maps to a shared embedding space and that the correlation between two modalities can be used as a measure of similarity between the two modalities. Therefore, the graph structure improves dynamically during training, purely as a result of the minimization of the contrastive loss. We evaluate AutoBIND on a wide variety of datasets, including tasks such as Alzhiemer's disease detection and house price prediction, and across a broad range of data modalities: 3D images, 2D images, and tables. We also show that AutoBIND outperforms previous methods on these tasks, highlighting the generalizablility of the approach.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper proposes AutoBIND, a new multimodal contrastive learning framework that can learn representations from any number of modalities without needing hand-crafted architectures. AutoBIND uses a graph-based approach to automatically select the most correlated modalities for contrastive learning. This allows it to be robust to missing modalities, as the graph structure is dynamically updated during training. Experiments across diverse datasets and modalities demonstrate AutoBIND's ability to generalize and superior performance compared with previous approaches.
+
+### Strengths
+1. The paper proposes a reasonable solution to an important problem. 
+
+2. The experimental results demonstrated satisfactory performance in a number of mixed-modality problem settings. 
+
+3. Visualization and interpretation of the graph structure formed during the learning are interesting.
+
+### Weaknesses
+1. The overall methodology is a simple application contrastive learning to multiple modalities. It is not clear why it outperforms previous approaches. Is it due to the additional modalities or due to one or several components in the proposed method? 
+
+2. The MMST is constructed in every epoch (and also pruned)? If so, will this process converge to a stable tree and how does the final tree depend on the initial encoding quality (which determine the initial tree). 
+
+3. The paper has the word "optimal" in the title, but there is no discussion or proof on the optimality. 
+
+4. Eq (3) is not explained clearly. Is it a theorem?
+
+### Questions
+See weaknesses.
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+1 poor
+
+---
+
+## Human Reviewer 2
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.
+
+### Summary
+This paper proposes a method to learn multimodal representation using a graph representation. The target representation is per modality, i.e. each modality should yield a separate vector of a fixed size (=128). Each node in the graph represents a separate modality. A graph is constructed in an order to maximize the correlations between modalities iteratively. The method was evaluated on two multimodal datasets for prediction tasks.
+
+### Strengths
+A graph construction method was used and the resulting graph may be used to understand the internal correlation of data.
+The proposed method is very simple to implement.
+
+### Weaknesses
+Technical descriptions are very unclear. Mathematical formalization is inconsistent and doesn't make sense. The algorithm description is incomplete. I think I can still guess what the paper tries to do, but not because the presentation is good, but because the method is extremely simple. 
+
+At high level, there is no proof or evaluation showing that this method is "optimal" or "generalizable". There is no proof that the proposed algorithm can optimize the global objective (eq 1). 
+
+Eq (1) is the global objective by choosing optimal {f_i}. z_i is instance Z_i should be the set. In the first equation in Sec 2.1 (doesn't have a numbering), the papers uses dot product/norms of set (Z). This must be z_i. But then w_ij should be defined in the whole set, so probably needs summation. 
+
+The next paragraph introduces d_ij, which is just 1/Sim(). There is no point of defining this term because this term never appears in the rest of the paper. 
+
+Eq (2) and (3) make no sense. First, i and j now represent sets, not instances. Eq (2) is only valid when the inputs are two different modalities. Eq (3) is passing same modality in each of the loss terms. 
+
+Algorithm 1 also uses in consistent and undefined terms like "correlation factor" "get_embedding". How do you train the contrastive model on modality pair? It's never explained, although I guess it's probably done using eq (2). 
+
+The evaluation is another (bigger) limitation of the current paper. This paper seeks to learn a "unimodal" representation that can be learned from multimodal sources. This representation takes one modality as an input and produces a representation of fixed size. This can be useful in many situations, like when there is only one modality present per instance at test time. This paper doesn't consider any situation like this in evaluation. Instead, it uses all the modalities and concatenates all the representations to get a *joint* representation. And of course, this leads to a better prediction performance than using a single modality, but this is pointless. Maybe Table 3 tried to evaluate the model when each modality is separately used. But Table 3 is never referred in the main body and there is not much explanation. More importantly, if Table 3 was indeed from the proposed method, this needs to be compared with some references (which learn multimodal representation like this paper). Currently this table delivers no information. 
+
+There is an arbitrary step of pruning nodes, but there's no ablation study showing its impact. So I assume that the authors needed to do this just to improve the performance. It was argued that the current method is robust against missing variables, but again there was no experiment designed to verify. 
+
+Overall, I think the current paper needs a lot more improvement before it can be published.
+
+### Questions
+Please see above and correct me if I was mistaken.
+
+### Soundness
+2 fair
+
+### Presentation
+2 fair
+
+### Contribution
+1 poor
+
+---
+
+## Human Reviewer 3
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper proposes a method name AutoBIND, which construct a graph on different modalities and dynamically update the graphs using a minimum spanning tree algorithm. This paper evaluates AutoBIND on a wide variety of datasets and show that AutoBIND outperforms previous methods.
+
+### Strengths
+1. This paper Introduces AutoBIND, a novel framework that efficiently handles multiple modalities and is robust to missing modalities through a dynamic graph-based approach. Utilizing graphs to learn multimodal representations is reasonable, as this structured data can effectively model the relationships between different modalities.
+2. Extensive Experiments across various datasets demonstrate AutoBIND's effectiveness in multimodal contrastive learning tasks.
+
+### Weaknesses
+1. In section 2.1, the definition of \(d_{ij}\) appears problematic. How is the value determined when the cosine similarity \(Sim(Z_i, Z_j)\) is zero? Moreover, when \(Sim(Z_i, Z_j)\) approaches 0, \(d_{ij}\) can become extremely large (if positive) or extremely small (if negative). This definition seems to be flawed at a fundamental level.
+2. What is the specific definitions of correlated and uncorrelated modalities?
+3. The symbol system is a mess. i,j are modality indices in equation 1 but are sets in equation 3, making this paper is hard to understand. 
+4. Overall, the paper lacks a clear structure and appears to be written informally.
+
+### Questions
+NA
+
+### Soundness
+1 poor
+
+### Presentation
+1 poor
+
+### Contribution
+1 poor
+
+---
+
+## Human Reviewer 4
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper introduces AutoBIND, a contrastive learning framework that can learn representations from an arbitrary number of modalities. 
+This approach leverages data from heterogeneous sources such as images and text and is able to deal with missing modalities as it can dynamically update the graph structure during training. 
+AutoBIND uses a graph-based approach to automatically select the most correlated modalities from fully connected graph/MST and a contrastive loss to learn the representations. 
+The method is evaluated on Alzheimer's disease detection and house price prediction with 3D images, 2D images, and table modalities and outperforms existing baselines.
+
+### Strengths
+The paper motivates well by tackling one of the most important problems in multimodality learning, which is dealing with a numerous number of modalities and missing modalities.
+The author proposes a simple yet effective method for extending the ImageBind pairwise contrastive learning with graph learning to learn the relation between the modalities. Results demonstrate decent results in the two downstream datasets, showing the effectiveness and interpretability of the method.
+
+### Weaknesses
+(1) Although the paper claims to bind multimodalities, in practice, it deals with only three modalities (image, text, tabular) and their corresponding features (various columns in the data). Such a number of modalities is relatively common in multimodal learning, unlike current works that learn from up to five modalities that contain image, video, text, audio, and IMU.
+(2) The modalities claimed in the current setting are more like different features in the same modalities (bedroom, bathroom, etc ...). Hence, the missing modality setting is more like missing features, which is a relatively explored problem. 
+(3) It will be better to include the 
+ImageBIND method as the baseline in Table 1, to confirm if it is the proposed method or the use of MM features that achieve performance improvement.
+(4) Due to the loss design including all pairs of combination (or MST edges), it will be better to include a more rigorous run-time analysis (how the method scales in adding more modalities), computation resource used, and parameter used. It is very likely that the performance gain was from more parameters in the encoder.
+
+Some typos
+Alzhiemer's should be Alzheimer's in the title  
+4.2.1
+
+### Questions
+It is unclear to me what the message the author tries to convey of the epoch 123 vs. 62,63,64 in Figures 3 and 4.
+For me, all of the graphs make sense in their own way, which is based on different interpretability of such a task.
+Also, the change between 62,63,64 is also not clear in Figure 4. In 62 and 64, the label is dependent on various features. However, it is only correlated to two of them in 63. It will be better for the author to elaborate more on these findings.
+
+### Soundness
+2 fair
+
+### Presentation
+2 fair
+
+### Contribution
+2 fair

human_reviews/f4HohsyNEk.md

@@ -0,0 +1,211 @@
+# NeuManifold: Neural Watertight Manifold Reconstruction with Efficient and High-Quality Rendering Support
+
+- Decision: Reject
+- Scores: 5, 3, 3, 3
+
+## Abstract
+We present a method for generating high-quality watertight manifold meshes from multi-view input images.
+Existing volumetric rendering methods are robust in optimization but tend to generate noisy meshes 
+with poor topology. Differentiable rasterization-based methods can generate high-quality meshes  
+but are sensitive to initialization.  
+Our method combines the benefits of both worlds; we take the geometry initialization obtained from neural volumetric fields, and further optimize the geometry as well as a compact neural texture representation 
+with differentiable rasterizers. 
+Through extensive experiments, we demonstrate that 
+our method can generate accurate mesh reconstructions with faithful appearance that are 
+comparable to previous volume rendering methods while being an order of magnitude 
+faster in rendering. We also show that our generated mesh and neural texture 
+reconstruction is compatible with existing graphics pipelines and enables downstream 3D applications such as simulation. Project page: \url{https://sarahweiii.io/neumanifold/}.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper introduces NeuManifold, a method for creating watertight manifold meshes from multi-view images. The approach integrates neural field rendering with differentiable rasterization-based mesh reconstruction techniques, providing accurate reconstructions with authentic appearance. The contribution lies in introducing Differentiable Marching Cubes (DiffMC) to yield smoother surfaces, coupled with seamless integration with GLSL shaders for real-time rendering.
+
+### Strengths
+Originality: This method is a synthesis of various existing technologies. TensoRF provides initialization for the subsequent mesh optimization. Differentiable Marching Cubes further refine the mesh, and nvdiffrast is used for rendering. The entire process is amenable to gradient backpropagation.
+
+Quality: Experiments indicate an improvement in both appearance synthesis and mesh generation quality compared to previous research and existing methods.
+
+Clarity: The language used for explanation is generally concise.
+
+Significance: The differentiable Marching Cubes hold significance and can serve as inspiration for other related fields.
+
+### Weaknesses
+1. The majority of this method is a combination of previous works. DiffMC demonstrates a certain level of innovation, but the explanation of this part of the method is not sufficiently clear. For example, the specific implementation of the approach in this paper and its differences from the "Deep marching cubes: Learning explicit surface representations" paper are not discussed.
+
+2. The main text claims a 10x speedup in DiffMC compared to DMTet. However, this improvement is not intuitively and reasonably explained. Additionally, the speed comparison between DiffMC and DMTet should be reflected in the experimental section.
+
+3. The paper claims that the method can generate watertight and manifold meshes, but marching cubes do not necessarily guarantee manifold properties in regions with complex iso-surfaces. The authors should compare whether the proportion of manifold vertices generated in different scenarios has improved. Otherwise, this claim might be somewhat exaggerated.
+
+4. The specific details of how the finetuning is performed in the third stage are not well explained.
+
+### Questions
+1. What are the specific working details of DiffMC? Please provide more explanation.
+
+2. Can the proposed method completely guarantee the watertightness and manifold of the generated mesh?
+
+3. Why is the proposed method faster than DMTet? Please provide more explanation.
+
+### Soundness
+2 fair
+
+### Presentation
+3 good
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 2
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper proposes a geometry reconstruction method, based on the following: neural radiance fields, a differentiable form of marching cubes, and differentiable rasterization.  They leverage components from existing works such as TensoRF and nvdiffrast, in addition to implementing the differentiable MC part in CUDA for high performance.  The pipeline results in a high-quality, watertight manifold mesh, along with a "neural texture", that can be used for rendering and simulation.
+
+### Strengths
+The writing quality is generally good, and well-articulated.  The paper provides thorough evaluation on multiple datasets, and comparisons to multiple baseline methods.  The end-to-end differentiability of the system allows refinement of the initial geometry based on a rendering loss, which improves the quality of the reconstruction result.  The implementation and engineering, such as deploying the optimized model in GLSL shaders, is notable.
+
+### Weaknesses
+While the paper shows favorable results in terms of quality of reconstruction, it seems only about on par or marginally better than methods that it has been compared to.  It mostly connects components from existing works (TensoRF, nvdiffrast, Marching Cubes algorithm) into a well-engineered end-to-end pipeline, but I question whether the specific technical contributions of the paper are significant enough to stand on their own.
+
+The paper claims that it is the first differentiable Marching Cubes implementation, yet it's not clear whether this is indeed the case (see for example: Deep Marching Cubes, MeshSDF, Neural Marching Cubes, Flexicubes).  Furthermore, the differentiable formulation of MC presented in the paper is compared only to DMTet, but not to any other methods that might be even closer in spirit (i.e. those mentioned in the previous remark).  Therefore, it is hard to evaluate the level of contribution.
+
+### Questions
+- The paper strongly emphasizes the watertight manifoldness of the resulting meshes, but does not go into much detail about the specific technical contribution of this paper that results in this property.  Could you clarify whether this is indeed a contribution of this paper, or is it inherited from the other methods that the paper builds off of?
+
+- On page 4, it is stated that during the second, mesh optimization step, the *topology* is also optimized.  Could you elaborate/clarify what you mean by this exactly?
+
+- On page 5: "given that most real-world objects tend to be axis-aligned" -- could you briefly explain why this is a reasonable assertion to make?
+
+- SDF-based reconstruction methods (e.g. NeuS) claim that the raw density field is optimized only to satisfy the rendering objective (and is also biased away from the true surface), and therefore are not ideal for surface reconstruction.  In this paper, you argue that this is unnecessary and that using the density field directly is a better choice.  Could you elaborate on this choice?
+
+### Soundness
+2 fair
+
+### Presentation
+3 good
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 3
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The manuscript presents a differentiable pipeline for generating watertight manifold meshes from multi-view images observation. The NeuManifold leverages TenorRF [Chen et al, 2022a] to recover a density-based representation. The density is then converted to SDF and later meshes using differentiable marching cubes presented in FlexiCubes [Shen et al, 2023]. The final optimization is baed on nvdiffrast [Laine et al, 2020] is done for improving details. The performance of NeuManifold is compared with prior work on the public available NeRF-Synthetic dataset.
+
+### Strengths
+-	The paper is well-written and clear to follow. The figures are illustrative to showcase the pipelines and performance variations. 
+-	The present idea is intuitively sound (except for the first component), with each stage performing meaningful tasks. 
+-	The problem is of interest to the research community, as image-based watertight reconstruction can be widely applicable in many industries.
+
+### Weaknesses
+-	Theoretical Limitations
+  - The suggested approach is asserted to possess the capability to restore watertight manifold meshes. Nevertheless, these meshes are directly generated from density data with marching cubes, and as a result, there could be arbitrary noisy surfaces extracted, unlike more principled representations such as SDF. The absence of a theoretical framework to ensure a principle density formulation towards watertight mesh contradicts the claimed contributions in the manuscript.
+
+-	Lack of Innovation: 
+  - Although the overall workflow appears logically coherent, the suggested method primarily amalgamates various pre-existing approaches. The rendering networks draw inspiration from TensorRF [Chen et al, 2022a], the notion of differentiable marching cubes originates from FlexiCubes [Shen et al, 2023], and the concept of differentiable rasterization is credited to nvdiffrast [Laine et al, 2020].
+
+
+-	Inadequate Experiment Results:
+  - The manuscript asserts that "SDF-based techniques often suffer from reduced visual fidelity and the loss of high-frequency geometric details" (Page 4, Section 3.3). However, the paper overlooks recent breakthroughs in SDF-based neural surface reconstruction algorithms capable of generating watertight, high-fidelity meshes from images. Numerous recent advancements, exemplified by Neuralangelo [Li et al, 2023], NeuS2 [Wang et al, 2023], and HF-NeuS [Wang et al, 2022], have emerged in this field. Without comparing against such approaches, it is difficult to understand how much progress NeuManifold has made. Furthermore, building NeuManifold on existing SDF-based approaches seems to be a more logical choice.
+  - Moreover, the current results are only constrained to synthetic scenes using NeRF-Synthetic. It will be interesting to see results from real-world captures.
+
+References: 
+
+Li, Zhaoshuo, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. "Neuralangelo: High-Fidelity Neural Surface Reconstruction." In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 8456-8465. 2023.
+
+Wang, Yiming, Qin Han, Marc Habermann, Kostas Daniilidis, Christian Theobalt, and Lingjie Liu. "Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction." In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 3295-3306. 2023.
+
+Wang, Yiqun, Ivan Skorokhodov, and Peter Wonka. "Hf-neus: Improved surface reconstruction using high-frequency details." Advances in Neural Information Processing Systems 35 (2022): 1966-1978.
+
+### Questions
+-	What are the theoretical novelties of the framework such that it is more suitable to the ICLR audience?
+- Can authors provide comparisons against the prior works using SDF-based surface reconstruction approaches, such as those listed above?
+-	How well does the pipeline work if the NeRF is based on recent SDF-based surface reconstruction methods? 
+-  How well does the approach work for real-world captures?
+-	The manuscript claims in Figure 5 that it is safe to assume axis alignment between marching cubes coordinates and real-world objects. Without explicit alignment, this may not be the case. Can the authors provide more details on this?
+-	Figure 3 compares marching cubes/tetrahedral results with and without “exp()” on SDF fields. Why is this important given SDF is already meaningful for surface extraction?
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+1 poor
+
+---
+
+## Human Reviewer 4
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.
+
+### Summary
+This paper presents a method that is able to generate high quality watertight manifold meshes from multi-view input images. The proposed method starts from a geometry initialized from neural volumetric fields and further optimizes the geometry and a texture representation with differentiable rasterizers. Experiments show that the generated mesh and the texture reconstruction are compatible with existing graphics pipelines.
+
+### Strengths
+1. The proposed method generates meshes that are applicable to simulation and geometry processing tasks.
+
+2. The proposed method can be integrated with shaders.
+
+### Weaknesses
+1. The proposed method shown in Figure 2 is a combination of existing papers. For example, volumetric rendering initialization, the geo & app networks and nvdiffrast are all from existing papers. Hence, the novelty of this paper is very limited.
+
+2. How is DiffMC compared to Deep Marching Cubes (https://github.com/yiyiliao/deep_marching_cubes), Neural Marching Cubes (https://github.com/czq142857/NMC), Flexicubes (https://research.nvidia.com/labs/toronto-ai/flexicubes/)? What is the advantage of DiffMC over these methods?
+
+3. While the method is able to generate watertight manifold meshes, it means that the images depicting a scene have some constraints. For example, if the images were depicting the scene of Figure 7 (c) where a blanket is draped onto a chair. I am curious about how the mesh the proposed method reconstructs looks like. The blanket should just be a flat sheet and it shouldn't be represented as a watertight mesh. Similarly, the leaves in Figure 7 (d) shouldn't be represented with watertight meshes.
+
+4. Is the method able to reconstruct objects of a scene into multiple mesh components? For example, in Figure 7 (b) the scene can at least be decomposed into 4 components, a plate, a hotdog and the sauces. Are we able to get four meshes by applying the proposed method to this scene? There are some works looking at understanding the semantics of the scene and perhaps that's good information that can be leveraged to do so.
+
+5. For some physics simulation tasks knowing the internal structure of the object is also important. Is the proposed method infer the internal structure of the reconstructed mesh?
+
+### Questions
+See the questions above.
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+2 fair

human_reviews/fpYIlzOpIA.md

@@ -0,0 +1,138 @@
+# NLPBench: Evaluating Large Language Models on Solving NLP Problems
+
+- Decision: Reject
+- Scores: 6, 5, 5
+
+## Abstract
+Recent developments in large language models (LLMs) have shown promise in enhancing the capabilities of natural language processing (NLP). Despite these successes, there remains a dearth of research dedicated to the NLP problem-solving abilities of LLMs. NLPBench includes questions with context, in which multiple sub-questions share the same public information, and diverse question types, including multiple choice, short answer, and math. Our evaluation, centered on LLMs such as GPT-3.5/4, PaLM-2, and LLAMA-2, incorporates advanced prompting strategies like the chain-of-thought (CoT) and tree-of-thought (ToT). Our study reveals that the effectiveness of the advanced prompting strategies can be inconsistent, occasionally damaging LLM performance, especially in smaller models like the LLAMA-2 (13b). Furthermore, our manual assessment illuminated specific shortcomings in LLMs' scientific problem-solving skills, with weaknesses in logical decomposition and reasoning notably affecting results.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper presents a benchmark dataset for NLP-related questions for LLMs to answer. The collected dataset, NLPBench, consists of 378 college-level NLP questions and various LLMs are tested on this benchmark. In addition, this paper also compares different prompting strategies, but finds out that the results are not consistent with the "advancement" of the technique.
+
+### Strengths
+* Interesting angle and scope for LLM evaluation
+* Experiments of the effectiveness of different prompting strategies under this task
+
+### Weaknesses
+* The scope of the evaluation is quite limited
+* The dataset is small and relatively the evaluation cost is high
+
+### Questions
+* Could this approach be adapted to other areas or domains? Would be nice to discuss about it and also the transfer cost.
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 2
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+The authors propose a new LLM benchmarking dataset consisting of 378 questions collected from Universities’ NLP final exams with short-answers,  multiple-choice questions, math questions with step-by-step solutions. Some questions come with a context information. Different LLMs (GPT-3.5/4, PaLM-2, and LLAMA-2) are benchmarked on the dataset with different prompting strategies with a combination of zero-shot, few-shot and chain of thought (CoT) and tree of thought (ToT). The experimental results show that prompting strategies such as can reduce performance, in particular for small size models. The evaluation shows LLM limitations in reasoning, logical decomposition.
+
+### Strengths
+- This work aim to expand the scope of LLM benchmarking and address new perspectives by contributing new datasets and benchmarking scenarios
+- The evaluation includes both close and open models. It is important to understand the gap between these model types
+
+### Weaknesses
+- The size of the dataset is relatively small and might limit the conclusion drawn from the results.
+- The gap filled by the proposed by introducing NLPBench (Table 6) is rather narrow and needs stronger justification. It covers a broader context than the claimed main focus of NLP-related topics such as Math. 
+- The analysis conclusion are mostly known, such as small LLMs have inconsistent results with advanced prompting. LLM limitation on problem-solving tasks.
+
+### Questions
+- How open questions have been evaluated ? The paper mentions only human annotation of errors.
+- Under Inaccessibility, details are missing. How did you make sure that the questions are not part of other datasets ? What are risks of contamination by LLM training data  ?
+- I expected that the top-3 models is including PalM but in 4.2, Llama-2 was taken instead. Is there a justification for this choice ?
+
+### Soundness
+2 fair
+
+### Presentation
+2 fair
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 3
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The authors develop a new benchmark dataset (NLPBench) to evaluate the large language models for solving NLP problems. NLPBench comprises 378 college-level NLP questions (with and without context) spanning various NLP topics sourced from some University's prior final exams. NLPBench was evaluated on different LLMs such as GPT-4, PaLM-2, and LLAMA-2 using advanced prompting strategies including chain-of-thought (CoT) and tree-of-thought (ToT) and different decoding strategies such as self-consistency. The results show that NLPBench illuminates specific shortcomings in LLMs’ scientific problem-solving skills, with weaknesses in logical decomposition and reasoning notably affecting performance.
+
+### Strengths
+1. Introduces a new dataset that challenges the current state-of-the-art prompting strategies and which is useful in evaluating the performance of LLMs.
+
+2. The paper is well-written and easy to follow.
+
+3. The authors carried out extensive experiments and evaluations that included recent prompting approaches and LLMs.
+
+### Weaknesses
+In general, NLPBench is useful, but I think there are some clarities on how it was collected that are missing:
+
+1. How did you get access to the final exams included in the dataset?  How do you ensure that these exams were not already online and that some recent models like GPT-4 have already included them in their training dataset? It was not clear from the paper how you checked that.  You mentioned that you curate questions that are not readily accessible online and couldn’t be easily extracted or transformed into text, but how do you measure that specific content cannot be easily extracted or transformed into text? what reliable techniques did you use to do that? 
+
+2. In what years these final exams were given? This information might help in the development of future related datasets. 
+
+3. In the data selection process, you mentioned that you selected 400 questions out of 1000 total questions, what were the criteria? Was it a random selection?
+
+4. The dataset size looks small, why in Table 6, was not there any comparison for the NLPBench size to other benchmarks?
+
+5. The very closely related benchmark (SciBench) also includes math problems, did you check if NLPBench's math-related problems are not also in SciBench? I did not see this evaluation in the paper.
+
+6. How many expert human annotators were involved in the dataset processing?
+
+**Minor:**
+Some typos: 
+- Page 3 in the paragraph that describes the complex structure: multi-tern -> multi-tern
+- Page 4: zero-shot and few-shot(FS+ToT, FS+ToT) -> (ZS+ToT, FS+ToT)
+
+### Questions
+- How many universities were included in the data collection?
+- For other questions, please check the weakness section.
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good

human_reviews/gcEhF4nuYI.md

@@ -0,0 +1,187 @@
+# FTP: A Fine-grained Token-wise Pruner for Large Language Models via Token Routing
+
+- Decision: Reject
+- Scores: 3, 3, 3, 3
+
+## Abstract
+Recently, large language models (LLMs) have demonstrated superior performance across various tasks by adhering to scaling laws, which significantly increase model size. However, the huge computation overhead during inference hinders the deployment in industrial applications. Many works leverage traditional compression approaches to boost model inference, but these always introduce additional training costs to restore the performance and the pruning results typically show noticeable performance drops compared to the original model when aiming for a specific level of acceleration. To address these issues, we propose a fine-grained token-wise pruning approach for the LLMs, which presents a learnable router to adaptively identify the less important tokens and skip them across model blocks to reduce computational cost during inference. To construct the router efficiently, we present a search-based sparsity scheduler for pruning sparsity allocation, a trainable router combined with our proposed four low-dimensional factors as input and three proposed losses. We conduct extensive experiments across different benchmarks on different LLMs to demonstrate the superiority of our method. Our approach achieves state-of-the-art (SOTA) pruning results, surpassing other existing pruning methods. For instance, our method outperforms  BlockPruner and ShortGPT by approximately 10 points on both LLaMA2-7B and Qwen1.5-7B in accuracy retention at comparable token sparsity levels.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+5
+
+### Summary
+In this paper the authors propose a method for accelerating inference in fine-tuned Transformer models by learning a routing function that selects whether to calculate updated representations for each token at each layer. The method works by first learning a “static” token router to identify fixed token positions to remove at each layer, with sparsity ($k$ for selecting top $k$ tokens) learned using a genetic algorithm approach. Given those target sparsities, and the static router, dynamic token-level routers are learned for each layer, using a lightweight set of features: position, absolute attention score, relative attention score rank, and layer target sparsity. The dynamic router is learned using standard SGD with a combination of losses incorporating the sparsity constraints, and distillation wrt the original model and static router. There is then an additional finetuning phase for the static sparsities using the learned dynamic router.
+
+In experiments on Llama-2-7B, Llama-2-13B, Llama-3-8B, and Qwen1.5-7B fine-tuned on the Alpaca dataset and evaluated on HellaSwag, MMLU, ARC-easy, ARC-challenge, and WinoGrande, the authors show that the approach can achieve better end-task performance at a comparable compression rate (sparsity level) compared to previous work.
+
+### Strengths
+- **Important and practical problem:** Improving LLM inference efficiency is an important area of research with many practical applications.
+- **Simple and intuitive solution based on analysis:** The solution is intuitive and includes interesting innovations, such as the simple classifier for learning dynamic token routing. It’s based on an initial analysis which demonstrates token redundancy across two model families.
+
+### Weaknesses
+I really wanted to like this paper, I think it's an innovative approach, but as the paper is currently written the results simply do not support the claims, due to highly insufficient comparison with prior work, particularly as it relates to computational efficiency.
+
+- **Paper could be written more clearly:** In the abstract and introduction, the approach is motivated in part by emphasizing that the approach does not require re-training, however, training is in fact required, to learn the sparsity levels and routing function (About 3 hours of training on a 64GB AMD MI250 GPU). This should be clarified throughout the paper, and it should be made clear early on that this is a task-specific approach for finetuning models. Additionally, various details are missing or poorly located throughout the paper. For example, “token length” is not clearly defined, training time for the dynamic router is included in the main paper but for the static router, and when the static router training is reported in the appendix, it’s not clear whether that includes the final finetuning phase. Batch size is not reported for inference experiments, which is critical for determining the practicality of the approach on accelerator hardware. If the approach only works with batch size 1, then this needs to be clearly addressed and motivated.
+- **Related work is insufficiently discussed:** The related work section could more clearly contrast with previous work, and some related work is missing. For example, foundational work on dynamic token-wise computation in Transformers [(Dehghani et al. 2019)](https://arxiv.org/abs/1807.03819), recent work on dynamic token-level routing [(Ainslie et al. 2023)](https://arxiv.org/abs/2303.09752) as well as work on token pruning in BERT models, e.g. [(Goyal et al. 2020)](https://arxiv.org/abs/2001.08950), [(Dai et al. 2020)](https://arxiv.org/abs/2006.03236), and others (these can be used as a starting point), which makes me concerned that other important related works might be missing. The MoD model is highly related, since it also does token-level dynamic routing, but its relation to the proposed work is barely discussed, with no motivation for why different design choices are made in the proposed approach, and the authors don’t compare to this highly related work in experimental results. 
+- **Insufficient efficiency evaluation compared to previous work, unclear whether approach leads to practical improvements:** Table 6 reports inference efficiency acceleration of the proposed model at different sparsity levels and across different maximum sequence lengths, but does not compare with any previous work in terms of efficiency. Sparsity levels for different approaches are reported in the main results table, but those are insufficient for determining whether there are practical efficiency improvements. Further, while the training efficiency of the method is claimed as a key contribution of the work, there are no experiments comparing training efficiency to prior work. 
+- **Minimal inference efficiency acceleration:** The maximum inference efficiency acceleration reported in Table 6, for max sequence lengths of 2k, is 1.6x. It’s unclear whether this improvement is achievable with batching; prior work with less dynamic resource allocation does so because it is necessary to leverage the practical computational improvements of batching. This needs to be addressed.
+
+### Questions
+- How does the tradeoff between inference efficiency and end-task performance compare to baselines? It would be good to see a plot with end-task accuracy on one axis and latency/speedup on the other axis, to see how this approach compares in terms of Pareto optimality to existing approaches. 
+- In Table 6, what are the actual average token lengths for each set of token length experiments? What batch size is used?
+- Is the approach compatible with acceleration using batching?
+- Is token length referring to total input/output sequence length, or just input sequence?
+
+Notes:
+- Figure 1 could be a bit clearer. I like the color coding for compute vs. skip, and if you continued those colors for the arrows in the digram (for example where you currently use dotted vs. solid arrows), I think this would help. Then, I think you could collapse the two versions of the “token router” zoom in on the right into one, and merge the “token router” component into the circle that determines whether to compute or skip (rather than having additional arrows between the token router and the circle) — the circle seems redundant, since is that not just exactly the token router component? 
+- Minor grammatical errors throughout the paper, I recommend using Grammarly, an LLM or similar tool to proofread and edit for grammar.
+- Fix latex quotation marks on line 741.
+- Add information on training time for static router to main paper, and include time for the final finetuning of the static router after training the dynamic router (this would happen naturally with a more detailed comparison of training time to prior work.)
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 2
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+This paper proposes a fine-grained token-wise pruning approach (FTP) for large language models (LLMs) to address the issue of high computational overhead during inference. The method involves a learnable router that adaptively skips less important tokens across model blocks. The authors present a search-based sparsity scheduler, a trainable router with specific input factors and losses, and conduct extensive experiments to demonstrate its superiority over existing pruning methods.
+
+### Strengths
+The proposed FTP method is a novel contribution to the field of LLM pruning. The idea of using a learnable router to adaptively identify and skip less important tokens is intuitive and has the potential to significantly reduce computational cost during inference without sacrificing too much performance.
+
+The authors conduct extensive experiments on various LLMs and benchmarks, comparing their method with several state-of-the-art pruning techniques. The results clearly demonstrate the superiority of the proposed method in terms of accuracy retention and sparsity levels, which adds credibility to the research.
+
+### Weaknesses
+Incomplete comparison of inference speedup. In Table 6, only the inference speedup ratio of the current model is presented, lacking a comparison with other baseline models under the same conditions. It is not clear whether the FTP's token selection method truly brings significant inference acceleration compared to other methods.
+
+Limited exploration of acceleration ratio across different lengths. There is no comparison of the acceleration ratio with other models at different token lengths. It remains uncertain whether the inference speed improvement of this model becomes less significant as the sentence length increases.
+
+### Questions
+-In Table 6, why was only the inference speedup of the current model presented without a comparison with other baseline models? How can we be sure that the FTP's token selection method effectively accelerates inference compared to other approaches?
+
+-Have you considered comparing the acceleration ratio of this model with other models at different token lengths? If so, what are the results? Does the inference speed improvement of this model decrease as the sentence length grows?
+
+-Have you considered integrating the Router and the Sparsity Scheduler? If so, what are the potential benefits and challenges of such an integration?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+---
+
+## Human Reviewer 3
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+This paper proposes a fine-grained token-wise pruning method, using a learnable router to adaptively skip less important tokens during inference, reducing computational overhead. The approach includes a search-based sparsity scheduler to construct the router efficiently. Experiments demonstrate that the proposed method achieves state-of-the-art pruning results.
+
+### Strengths
+1. The story is concise and easy to understand.
+2. The method seems concise yet effective.
+
+### Weaknesses
+Although the approach seems concise yet effective, the lack of many details makes the experiments difficult to find convincing, and some statements are also confusing.
+1. In lines 14–16 of the abstract and line 53 of the introduction, I would assume that your approach overcomes previous issues without needing any training. However, training is still required, which is somewhat confusing.
+2. Your experiment’s comparisons seem unfair. On one hand, the baselines such as LaCo and LLMPruner, do not require any training but your method needs training. On the other hand, it seems that your compression ratio is the proportion of skipped token hiddenstates, while the baselines pruning methods calculate the ratio based on the parameters removed, which eliminates all computation related to these parameters. Therefore, their actual computational overhead seems to be lower than yours. You could add results of the actual computation cost.
+3. Your training details are unclear. The distillation loss and guide loss are not formally defined, making it difficult to understand your training process. Additionally, the size of your training dataset in terms of token count should be provided.
+4. The difference between your method and hiddenstate sparsification is unclear. Although lines 137–140 seem to attempt an explanation, your approach essentially appears to be skipping hidden states.
+5. You should report on memory usage, as the additional router and intermediate variables introduce new overhead.
+
+### Questions
+1. Can the scheduler determine the sparsity rate needed for each block based on a given overall sparsity rate?
+2. Currently, in Transformer-based LLMs, the output of each layer is calculated like:
+attn_output = torch.matmul(attn_weights, value_states)
+which computes the hidden states of an entire sentence at once and how you skip the computation for certain tokens? Do you have any code available to share?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 4
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+5
+
+### Summary
+This study proposes a fine-grained token-wise pruning approach for the LLMs, which presents a learnable router to adaptively identify the less essential tokens and skip them across model blocks to reduce computational cost during inference. Comparative experiments were conducted on different benchmarks, performing better than several existing works.
+
+### Strengths
+1. This paper is well-organized and easy to follow.
+2. Comparative experiments were conducted on different benchmarks, performing better than several existing works.
+
+### Weaknesses
+1. The scientific issues of the proposed paper are unclear. Existing prune methods of LLMs aim to remove the redundant module. This paper fails to discuss the shortcomings or deficiencies of existing work and cannot present the paper's main contribution.
+2. The novelty of this paper is limited. Several works have discussed the Redundant issue, e.g., [1], [2], which cannot be considered one of the paper's main contributions.
+3. Pruning aims to remove unimportant parts, not redundant parts. High redundancy comes from high similarity, which is highly related to calculating multi-head attention (attention scoring comes from similarity). Obviously, removing these tokens affects the subsequent multi-head attention and FFN calculations.
+4. The authors spent much space describing the multi-head attention mechanism in Section 3.1, which has nothing to do with the author's tokenwise prune. All symbols in Eq. (1) and (2) no longer appear or are even mentioned in the following content. Since the page is limited, presenting more in-depth experimental content is more meaningful.
+5. Equal balance co-efficiencies in Eq. (6) result in pruning losses exceeding task training losses, leading the model to tend towards pruning rather than achieving good inference performance.
+6. The experiment is too routine to demonstrate the effectiveness and rationale of the proposed dynamic router.
+7. The authors have released no code. The implementation and credibility of the proposed method are still unclear.
+
+[1] Jun Kong et al., "Accelerating Inference for Pretrained Language Models by UnifiedMulti-Perspective Early Exiting," COLING-2022.
+[2] Josef Kalfka et al., "Spying on your neighbors: Fine-grained probing of contextual embeddings for information about surrounding words." ACL-2020.
+
+### Questions
+See Weaknesses
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2

human_reviews/gzqrANCF4g.md

@@ -0,0 +1,134 @@
+# Language Model Beats Diffusion - Tokenizer is key to visual generation
+
+- Decision: Accept
+- Scores: 8, 8, 8
+
+## Abstract
+\vspace{-3mm}
+While Large Language Models (LLMs) are the dominant models for generative tasks in language, they do not perform as well as diffusion models on image and video generation.
+To effectively use LLMs for visual generation, one crucial component is the visual tokenizer that maps pixel-space inputs to discrete tokens appropriate for LLM learning.
+In this paper, we introduce \modelname{}, a video tokenizer designed to generate concise and expressive tokens for both videos and images using a common token vocabulary.
+Equipped with this new tokenizer, we show that LLMs outperform diffusion models on standard image and video generation benchmarks including ImageNet and Kinetics. 
+In addition, we demonstrate that our tokenizer surpasses the previously top-performing video tokenizer on two more tasks: (1) video compression comparable to the next-generation video codec (VVC) according to human evaluations, and (2) learning effective representations for action recognition tasks.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+The paper proposes a novel visual tokenizer based on lookup-free quantization (LFQ). With the growth of the vocabulary size LFQ consistently improve both reconstruction and generation quality, which is in stark contrast with Vector Quantization (VQ) where an increased vocabulary size reduces reconstruction error but hurts generation results. The tokenizer can be integrated with MAGVIT and achieves state-of-the-art performance on video generation. The tokenizer can also improve video compression and video recognition.
+
+### Strengths
+1. Starting from an interesting finding that enlarging the vocabulary improves reconstruction quality but hurts generation results, the paper proposes solutions (LFQ) to tame both reconstruction and generation simultaneously.
+
+2. A detailed study of architecture modifications that improves up MAGVIT supported by extensive ablations. 
+
+3. The tokenizer is proved to benefit video generation, compression and recognition. It will potentially have a huge impact on the general audience of video understanding.
+
+### Weaknesses
+1. For video compression results, it would be better if there is a PSNR/LPIPS/MS-SSIM-bpp curve comparing the performance across different bpps.
+
+2. In the video recognition setup, it seems unnecessary to detokenize the visual tokens back to pixels since BEVT and BEIT can work with tokenized input. I understand one of the main reasons is that the underlying recognition model is the ViViT which takes raw pixels as input (as stated in the draft). However, you may also have a comparison with BEVT.
+
+### Questions
+See weakness.
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+4 excellent
+
+---
+
+## Human Reviewer 2
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper focuses on learning a video / image tokenizer to discretize video / images so that they can be modeled using a Language Model. They specifically introduce one innovation in this setting: A lookup free quantizer. They show that in this limit of using a large vocabulary and no lookup, the tokenizer reconstruction and LM generation quality both increase with vocabulary size. The authors also show that the learned tokenizer performs very well as a compression scheme.
+
+### Strengths
+The paper does a good job of motivating the core thesis of the paper: How to design a tokenizer for image / video? The authors also do a good job of presenting this idea to the uninitiated. It is also pretty clear that the work is significant to academia and industry given that it helps unify image generation with image understanding and natural language generation and understanding techniques. The ideas in the paper are well explained. The authors also do a thorough job of running experiments to substantiate many claims including numerous ablations. Some of the important technical insights like the lookup free quantizer (and in general lower dimensional code words) helping in generation quality are substantiated by experimental results
+
+### Weaknesses
+My main concern is the completeness of the exposition in the paper. The authors assume that the reader is familiar with the state of the art in video tokenization and details do get rather buried in the many “deltas” relative to the baseline. I do understand the space limitations but it might be helpful if the authors try to make the core system / model design more explicit. Lot of the ideas like factorization of the output space in the decoder (and associated weight tying) for example are just mentioned in passing.
+
+### Questions
+* The question “Why masked LM and not AR LM for image / video generation?” for evaluating the tokenizer was not clearly answered.
+* No explicit definition of the objective for training the tokenizer (loss function)
+* No mention of decoder in VQ-VAE and VQ-VAE loss used when we use no lookup
+* More motivation needed on why the authors choose to use a causal encoder for tokenizer when doing masked LM for image / video generation
+* It’s not clear why the authors tackle video generation if the aim was to understand the fundamentals of tokenization. It may be desirable for them to clearly motivate why they study videos and not images alone?
+* It may be interesting for the reader to understand the computational complexity of both the tokenizer (encoder) and the detokenizer (decoder) and how they compare with video or audio codecs
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+4 excellent
+
+---
+
+## Human Reviewer 3
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper introduces a novel visual tokenizer designed to enhance large language models in producing high-quality images and videos. Experimental results show that, when integrated with the proposed tokenizer, LLM surpasses diffusion models in standard benchmarks such as ImageNet, UCF-101, and Kinetics-600. Additionally, the paper presents promising results in video compression and representation learning.
+
+### Strengths
+1.	This paper presents the first evidence of large language models surpassing diffusion models on the ImageNet benchmark.
+2.	The paper proposes a novel lookup-free quantization approach, providing a promising direction for expanding vocabulary size in LLM-based visual generation.
+3.	The motivation is clear, and the overall presentation is coherent and easy to follow.
+4.	Good results on visual generation video compression, and video representation learning.
+
+### Weaknesses
+1.	While the presented method is tailored for masked LM, many of the prevailing and powerful LLMs, such as LLaMA [A], employ an autoregressive approach. Incorporating results from AR-LM would greatly enhance the paper's relevance to the community.
+2.	In Table 4, despite the good action recognition performance showcased by the proposed method, it doesn't conclusively establish its efficacy as a viable self-supervised pre-training target. Notably, some pivotal baselines, like pixel colors and the image descriptor from MaskFeat [B], are missing.
+
+[A] Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M. A., Lacroix, T., ... & Lample, G. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.
+
+[B] Wei, C., Fan, H., Xie, S., Wu, C. Y., Yuille, A., & Feichtenhofer, C. Masked feature prediction for self-supervised visual pre-training. In CVPR 2022.
+
+### Questions
+1.	In Figure 1, it's highlighted that the VQ generation FID sees a pivotal change at a vocabulary size of 2^14, while the LFG generation FID consistently improves. I'm curious to understand how the LFG generation FID would respond to even larger vocabulary sizes.
+2.	Regarding Table 3, what could be the reason behind the proposed method's PSNR and MS-SSIM values being inferior to those of the standard video codec?
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+4 excellent

human_reviews/h0ZfDIrj7T.md

@@ -0,0 +1,199 @@
+# Mixture-of-Agents Enhances Large Language Model Capabilities
+
+- Decision: Accept
+- Scores: 6, 8, 8, 8
+
+## Abstract
+Recent advances in large language models (LLMs) demonstrate substantial capabilities in natural language understanding and generation tasks. 
+With the growing number of LLMs, how to harness the collective expertise of multiple LLMs is an exciting open direction.
+Toward this goal, we propose a new approach that leverages the collective strengths of multiple LLMs through a Mixture-of-Agents (MoA) methodology.
+In our approach, we construct a layered MoA architecture wherein each layer comprises multiple LLM agents.
+Each agent takes all the outputs from agents in the previous layer as auxiliary information in generating its response.   
+MoA models achieves state-of-art performance on  AlpacaEval 2.0, MT-Bench and FLASK, surpassing GPT-4 Omni. For example, our MoA using only open-source LLMs is the leader of AlpacaEval 2.0 by a substantial gap, achieving a score of 65.1\% compared to 57.5\% by GPT-4 Omni.
+}
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4
+
+### Summary
+This paper proposes a mixture-of-agents (MoA) framework that leverages the collective insights of multiple large language models (LLMs).
+
+The core process involves several iterative steps: (1) multiple LLMs generate initial draft proposals; (2) these proposals are shared with all LLMs in the following iterations to generate refined proposals; (3) this second step is repeated multiple times, with an aggregator ultimately producing the final response.
+
+Experimental results on four benchmarks (AlpacaEval 2.0, MT-Bench, Arean Hard, and FLASK) demonstrate that MoA can outperform GPT-4 Omni.
+
+### Strengths
+1. To the best of my knowledge, the proposed framework is both novel and reasonable. MoA can be viewed as a specific method for combining multiple weaker models to create a stronger model.
+
+2. The model's performance is competitive, showing improvements over GPT-4 Omni on three benchmarks.
+
+### Weaknesses
+1.  The proposed model is more resource-intensive than single LLM-based models.  
+2.  Most evaluations use LC metrics, with only a limited evaluation on MATH tasks included in the appendix. Further evaluations on diverse tasks are necessary to illustrate the general advantages of the proposed method.  
+
+3. An important question is how to select the set of proposal LLMs. Currently, the paper demonstrates two setups: one with relatively large models and one with smaller models. However, there is no guideline or formal strategy for choosing LLM sets. Additionally, it would be interesting to see the performance of a mixed set containing a small number of large LLMs and a majority of small LLMs.  Combining models with different language data is also a very interesting topic. Another crucial question is determining the number of layers. Aside from treating it as a hyperparameter, is there a way to determine an optimal layer count for a given set of proposal LLMs?  
+
+4. GPT-4o is only used as the aggregator. It would be informative to evaluate its performance if it were included as one of the proposal LLMs.
+
+### Questions
+1.  Why do you call your method as mixture of agents instead of mixture of LLMs.   I did not see any behaviours of agents. 
+
+2.  Are there any lines missing in Figure 6 (a) and (b)?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 2
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4
+
+### Summary
+The paper proposes Mixture-of-Agents (MoA) as a possible pathway to leverage the collective strengths of multiple LLMs.  By stacking LLMs into multiple layers and employing the output of LLM in previous layers, the frameworks allows a iterative refinement and strike a balance between inference cost and the performance.
+
+### Strengths
++ Proposal of a new effective framework to employ the collective intelligence of multiple LLMs.
++ Empirical evaluation on AlpacaEval 2.0, Arena-Hard, and MT-Bench verifies the effectiveness of the proposed solution.
+
+### Weaknesses
++ Stacking LLMs into layers and revising the output obtained from previous layers seems like another form of model ensemble and I would suggest including model ensemble as one of the comparative methods.
++ In Figure 6, the max number of tflops among proposers in each MoA layer is used as an approximation of the total tflops of the entire layers since different proposers can run in a parallel way. However, the approximation is only reasonable when considering the inference latency for a single query. Otherwise, if we have a batch of data, with the same computation resources MoA can introduce extra latency since the parallel of proposers limits the data parallel size. 
++ The evaluation of MoA is only performed on instruction-following benchmarks. More evaluation of reasoning and commonsense knowledge benchmarks such as MMLU and BBH is necessary for a comprehensive understanding of the efficacy of MoA.   
++ Detailed ability on the specific role of each proposer is missing. How does each proposer contribute to the final output? Is it possible for a proposer to simply repeat the output from previous layers? Will the models in later layers revise the mistakes made in previous layers?
+
+### Questions
+See the weaknesses above.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+---
+
+## Human Reviewer 3
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4
+
+### Summary
+This work presents a Mixture of Agents (MoA) Architecture. This multi-layer architecture harnesses the collaborativeness of LLMs, where agents (models) in later layers aggregate and combine auxiliary information from previous layers with the user task to yield the final output. They evaluate this proposed architecture on AlpacaEval 2.0, Arena-Hard, MT-bench, and Flask and show that their architecture performance surpasses GPT-4o. 
+
+The "collaborativeness" of LLMs is defined as the phenomenon where the LLM generates better responses when presented with outputs from other models, even if these models are less capable. 
+
+The architecture presented has two types of models: "proposer models," which generate useful references for use by other models; these models offer "perspective" or might even act as auxiliary/additional information providers to the aggregator. The other type of models present are "aggregator models," which are proficient in combining responses from other models into a single high-quality output. 
+
+They not only show an MoA architecture but also a MoA-Lite architecture that outperforms GPT-4o and is both better and more cost-effective. They conduct various empirical experiments showcasing the effectiveness of their approach to MoA outperforming LLM-Ranking, the ability to incorporate best answers, and the effects of model diversity and different proposers.
+
+### Strengths
+- While collaborativeness has been harnessed in various ways, a layered funnel architecture in which earlier layers add information for later layers to consume and interplay to summarize these outputs efficiently to yield a final output has not been explored. 
+
+- The authors also thoroughly conducted their experiments to establish collaborativeness and benchmark various datasets. The usage of open-source models to showcase the results helps to make replicating these results possible. 
+
+- They also analyze the cost-effectiveness of these models compared to some base models, showing the effectiveness of applying this architecture to obtain better results at a cheaper rate.
+
+### Weaknesses
+- The idea of collaborativeness or hierarchical processing in LLMs is not exactly novel [1][2]; if you think of different layers in the architecture using the same model, this reduces to some form of iterative refinement of outputs as shown in [2]. 
+
+- Some of the analysis in the paper to support budget analysis is unclear.
+
+
+
+#### *References*
+1] [2308.10848] AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors 
+
+2] LEGO: A Multi-agent Collaborative Framework with Role-playing and Iterative Feedback for Causality Explanation Generation - ACL Anthology
+
+### Questions
+1] What is the point of the tflops consumption analysis? And where are the numbers for these obtained from? Gpt-4o and Gpt-4-turbo are closed-source models, so it is unclear where those numbers are obtained.
+
+2] In a similar vein, can tflops really be a proxy for latency where one is used to perhaps look at real-time serving of models while the other looks at consumption? Especially when the system in question is not locally run but uses an inference endpoint (as mentioned in section 3.1)
+
+3] While you analyzed a model's propensity to be an aggregator or a proposer, did you consider combining models that might be good at different kinds of tasks (for example, math or science)?
+
+4] Was there any reason to include LLM reasoning in the related work when the work does not explicitly show gains on any reasoning dataset and the evaluation is mostly on instruction following/preference benchmarks?
+
+Last but not least, just a suggestion for the paper's readability: it would significantly enhance the analogy to MoE if there was an explicit mention that MoE uses routing/gating networks to sometimes choose between routing between single or multiple experts. They specialize in different aspects vs traditional networks where all input data is processed through every layer. There is a mention in the work that the MoA, from the specification of the architecture, consolidates the role of the gating network and expert network using a LLM. Still, it is not the case on some level because it is not choosing which experts it will route its output from, and the input does pass through all layers, the aggregator picks and chooses information from the proposer LLMs but it does not necessarily always just drop some inputs.
+
+### Soundness
+4
+
+### Presentation
+3
+
+### Contribution
+3
+
+---
+
+## Human Reviewer 4
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+3
+
+### Summary
+The paper proposes a method to use multiple LLMs as agents to improve the quality of generated responses to arbitrary queries. Agents are deployed in mixture-of-agent (MoA) layers. In the first MoA layer, multiple agents generate an answer to the input query independently. In subsequent layers, each agent in layer $T$ generates an answer to the query while also seeing the answers of all agents from layer $T-1$. At the last layer, a single aggregator LLM is used to combine all answers generated by the agents in the last MoA layer into a single answer. The proposed method shows state-of-the-art performance on multiple benchmarks compared to recent models from the GPT family including GPT4-omni, GPT4-turbo.
+
+### Strengths
+- The idea of using multiple LLMs as agents to generate answers and to aggregate these answers in a sequence of mixture-of-agents (MoA) layers is interesting.
+- Very good performance compared to many relevant baselines.
+- Experiments are solid and well conducted, and the ablation studies give important insight into the most relevant aspects of the proposed method.
+
+### Weaknesses
+- The main weakness, in my opinion, is the fact the authors did not use human evaluation at any parts of their pipeline. I understand that LLMs can perform very well (according to measures of correlation to humans) on multiple benchmarks, but at least one small-scale human evaluation could have been provided as a sanity-check.
+
+### Questions
+- Do you use humans in your evaluation in any part of your paper? It seems that all your tests use (language) models for evaluation. Did you conduct any (manual/human) analysis of the outputs generated by the models at any stage?
+- An experiment that could be more prominently contextualised is the case where there is one aggregator/proposer agent in each layer (numbers in Table 3, discussion in **Effect of model diversity and the number of proposers**. This is kind of a middle-case between the LLM-as-reranker baseline (which you clearly outperform) and your proposed idea in full generality. In the case of a single proposer, your win rate in Table 3 is 47.8% (so not better than the baseline). I think it would be very valuable to include this setting as a baseline also in all other experiments and not only those for AlpacaEval2.0, since this really shows when the proposed method starts to show improvements to the baselines.
+- The ablations are very informative and well conducted, and covers all the important points one would want to see in the main paper.
+- Across the paper, there are mentions to GPT4 Preview (Figure 5a), GPT-3.5-turbo-0125 (Figure 3), GPT4-turbo (Figure 6a), GPT4-omni (in multiple places). Could you please introduce all the models from the GPT family you used across your experiments in a single place with an unambiguous description of their model name if available? The paper is in general very well organised, this is however one of the few points where this organisation could be improved.
+- In Figures 6a and 6b, please include in your description why you have multiple circles of the same size/colour, and provide some context on what you mean by '1', '2', '3'. For someone who read the paper and reach the figures that is very clear, but someone looking at the figure without that context fresh may not understand it. Also, is it correct to assume that each of the large blue circles (MoA, multi-proposer, with 3 MoA layers) is an MoA model with different configurations (like different LLMs as final aggregators)? If that is the case, it may make sense to add some nuance to your discussion of the results. In practice, for someone with a new problem who want to use your paper's idea, do you assume by reusing your best-performing MoA, they will also get the best results on their new problem? In other words: should you not include also something like the average over the 6 large blue circles (or perhaps over the 3 best large blue circles)?
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+3

human_reviews/iX7eHHE5Tx.md

@@ -0,0 +1,169 @@
+# REMEDY: Recipe Merging Dynamics in Large Vision-Language Models
+
+- Decision: Accept
+- Scores: 6, 5, 6, 8
+
+## Abstract
+Model merging has emerged as a powerful technique for combining task-specific vision models into a unified and multi-functional model. Previous methods represented by task arithmetic, have demonstrated effectiveness and scalability in this domain. When large vision-language models (LVLMs) arise with model size scaling up, this design becomes challenging to fuse different instruction-tuned LVLMs for generalization enhancement. The large scale and multi-modal nature of LVLMs present unique obstacles, including constructing reusable and modular components to accommodate the multi-component architecture of LVLMs and the requirement for dynamic fusion based on multi-modal input tokens. To address these challenges, we propose the \textbf{RE}cipe \textbf{ME}rging \textbf{DY}namics (REMEDY) method, a scalable and flexible paradigm for model merging in LVLMs. We first define reusable modules termed \textit{recipes} including the projector and shallow LLM layers, enhancing visual-language understanding. Then, we introduce a modality-aware allocator dynamically generates weights in a one-shot manner based on input relevance to existing recipes, enabling efficient cross-modal knowledge integration. REMEDY thus offers an adaptive solution for LVLMs to tackle both seen (i.e., multi-task learning) and unseen (i.e., zero-shot generalization) tasks. Experimental results demonstrate that our method consistently improves performance on both seen and unseen tasks, underscoring the effectiveness of REMEDY in diverse multi-modal scenarios.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+3
+
+### Summary
+The paper introduces REMEDY (Recipe Merging Dynamics), a model merging framework tailored for large vision-language models (LVLMs). REMEDY leverages modular "recipes" and a dynamic, modality-aware allocator for effective cross-modal knowledge integration, enhancing multi-task learning and zero-shot generalization across diverse vision-language tasks.
+
+### Strengths
+1. The paper presents a unique "recipe" construction framework, creating modular, reusable components that facilitate task-specific adaptation and multi-task learning within large vision-language models.
+
+2.  Extensive experiments on multiple datasets demonstrate REMEDY’s efficacy, highlighting superior performance over baseline methods on multi-task and zero-shot generalization. 
+
+3. The study provides a thorough analysis of task similarity and allocation strategies, offering insights into how specific recipes contribute to various tasks.
+
+### Weaknesses
+1.  The process of constructing recipes heavily relies on experimental insights, such as which layers and modules to fine-tune. This empirical approach may limit REMEDY's scalability and adaptability to new tasks or models without extensive trial-and-error adjustments.
+
+2.  Although REMEDY introduces modularity, the recipe construction phase lacks dynamic or automated selection mechanisms to identify optimal modules or layers. 
+
+3. The paper does not extensively discuss the computational demands of the proposed approach, especially in the context of training and inference with the modality-aware allocator.
+
+### Questions
+As shown in weaknesses.
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+---
+
+## Human Reviewer 2
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4
+
+### Summary
+This paper proposes REMEDY, a new technique for merging large vision-language models (LVLMs). It addresses the challenges of scale and multi-modal input by creating reusable fine-tuned modules called "recipes" and dynamically combining them using a "modality-aware allocator." This allocator analyses the input image and text to generate weights for each recipe, enabling adaptive fusion and improved performance on both seen and unseen vision-language tasks.
+
+### Strengths
+1. The analysis of finetuning strategies is interesting. The paper shows that fine-tuning the projector and shallow layers of the LLM leads to better improvements.
+2. The motivation that uses the most relevant experts to better answer user questions is reasonable.
+
+### Weaknesses
+1. There is a typo in line 054: the vision encoder should be 0.3B.
+2. There are some redundant definitions in this paper and the "recipes" are actually MoE-based LoRA layers. The use of global and module-wise routing is not novel.
+3. The experiments include only 4 seen tasks and evaluation benchmarks. And I suspect that the proposed method can generalize well. It is expected that REMEDY surpasses the zero-shot setting on seen tasks since it is trained on these datasets. My main concern is the improvement on unseen tasks, which appears weak overall.
+4. TextCaps and TextVQA use the same images, so TextCaps is not an unseen task if TextVQA is already seen.
+5. The baseline is somewhat weak. There are stronger open-source baselines available, such as LLaVA-NeXT and LLaVA-OneVision.
+
+### Questions
+Please see the weaknesses.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+---
+
+## Human Reviewer 3
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+4
+
+### Summary
+This paper introduces REMEDY, a model merging technique for LVLMs that addresses the challenges of combining knowledge across multi-modal tasks. REMEDY explores components of LVLMs and propose recipes for learn specific tasks. It further propose a dynamic allocator that learns weights for different recipes. Experiments demonstrate the effectiveness of the proposed method.
+
+### Strengths
+1. The method propose a new approach for model merging for LVLMs, which is effective for enhancing the multi-task learning capabilities of LVLMs.
+2. The paper is well-writing and the experiments are comprehensive.
+
+### Weaknesses
+1. The work trains the model in few-shot way, e.g., on seen tasks and expected to generalize to unseen tasks. But current MLLMs are expected to have strong zero-shot capabilities. And from the comparison results of Table 2, REMEDY seems not have explicit performance improvement compared to the zero-shot way or simple average way. Considering the complexity it adds on the training and inference stage, it may limit the practical use case of it.
+2. Does the model needs to compute the weights for different recipes and then update the model weights for each inference time. If so, how about the extra inference time introduced and the comparison with the baseline?
+
+### Questions
+How about the performance of REMEDY compared to LLaVA fine-tuned on the seen tasks?
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+3
+
+---
+
+## Human Reviewer 4
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4
+
+### Summary
+This paper introduces a straightforward approach to addressing the multi-task learning challenge in Large Vision-Language Models (LVLMs) using model merging techniques. Given the large parameter scale of LVLMs, the authors first select layers to merge, referred to as 'recipes,' based on preliminary experiments. These layers include the projection layer and the shallow layers of the LVLM. An attention-based module is then trained to dynamically assign merging weights to the 'recipes' based on the input multi-modal token, allowing for adaptive weighting to handle the complexity of VL tasks. The proposed method demonstrates improvements on both learned tasks and zero-shot generalization across multiple datasets.
+
+### Strengths
+- **Clarity**: The paper is well-structured and easy to follow.
+- **Motivation**: As the size of the visual language model continues to grow, it becomes costly to maintain a separate model for each specific task. It makes sense to combine task-specific models into a unified model through model merging that can simultaneously solve different downstream tasks.
+- **Novelty**: This paper designs a merging strategy that takes into account the characteristics of multimodal tasks and the difference in fine-tuning paradigms between LVLM and traditional models. Meanwhile, this paper also analyzes the influence of each component in LVLM on fine-tuning performance, which is suggestive for the follow-up works.
+- **Results**: The experimental results across multiple datasets effectively highlight the potential and robustness of the method.
+
+### Weaknesses
+- **Detail in Weight Generation**: The paper does not explain how the proposed attention-based module is trained to generate the merging weights. Neither the main text nor the supplementary materials provide insights into the optimization objective used. While the authors may consider this aspect elementary, I believe a detailed explanation is necessary.
+- **Analysis of Computational Overhead**: The work in this paper is carefully designed when selecting an extended module, which may reduce the cost of calculating overhead and model parameters. However, the experiment lacks analysis of the training and inference overheads.
+- **Details Error**: The introduction states that the new challenges facing the LVLM merger will be analyzed from three aspects, but only two aspects are introduced below.
+
+### Questions
+I have some additional concerns:
+
+- **Evaluation Metric**: What does the Hscore metric indicate? Is the average accuracy calculated as a direct average of the accuracies on the different datasets or is it weighted by the number of samples in the dataset?
+
+- **Score Visualization Analysis**: In Figure 4, the weight distribution across different layers shows considerable variation, with some layers dominated by distinct 'recipes.' This phenomenon is intriguing, and I would appreciate further explanation from the authors. For VizWiz, an unseen dataset with the same VQA task, why does the "Recipe" of TextVQA have a smaller layer-wise score?
+
+Finally, the proposed Modality-aware Allocator is designed for multimodal tasks, which is one of the advantages over previous research on visual models. Therefore, I am curious whether using only visual or textual input as clues for allocation will result in a performance degradation?
+
+### Soundness
+4
+
+### Presentation
+3
+
+### Contribution
+3

human_reviews/juxbsQEuTZ.md

@@ -0,0 +1,174 @@
+# Sometimes I am a Tree: Data Drives Unstable Hierarchical Generalization
+
+- Decision: Reject
+- Scores: 5, 8, 3
+
+## Abstract
+Neural networks often favor shortcut heuristics based on surface-level patterns.
+As one example, language models (LMs) behave like n-gram models early in training. 
+However, to correctly apply grammatical rules, LMs must rely on hierarchical syntactic representations instead of n-grams.
+In this work, we use cases studies of English grammar to explore how latent structure in training data drives models toward improved out-of-distribution (OOD) generalization. 
+We then investigate how data composition can lead to inconsistent OOD behavior across random seeds and to unstable training dynamics. Our results show that models stabilize in their OOD behavior only when they fully commit to either a surface-level linear rule or a hierarchical rule. The hierarchical rule, furthermore, is induced by grammatically complex sequences with deep embedding structures, whereas the linear rule is induced by simpler sequences. 
+When the data contains a mix of simple and complex examples, potential rules compete; each independent training run either stabilizes by committing to a single rule or remains unstable in its OOD behavior. These conditions lead `stable seeds' to cluster around simple rules, forming bimodal performance distributions across seeds. We also identify an exception to the relationship between stability and generalization: models which memorize patterns from low-diversity training data can overfit stably, with different rules for memorized and unmemorized patterns.
+Our findings emphasize the critical role of training data in shaping generalization patterns and how competition between data subsets contributes to inconsistent generalization outcomes across random seeds.git}.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4
+
+### Summary
+This study analyzes the role of the syntactic complexity in language models’ acquisition of hierarchical inductive biases. Specifically, by generating sentences with varying syntax tree depths and of varying structures (some with mixed scoping and others with forward scoping), the features that lead to linear or hierarchical generalization are isolated. Additionally, by mixing these types of data, the authors conduct a grokking-style analysis of when stable, unstable, hierarchical, linear, or mixed generalizations are likely to be learned. 
+
+It is found that including sentences with a syntactic tree depth of at least 3 is required for hierarchical generalizations to be stably acquired, and that mixing depths in the training data can lead to unstable generalizations. It is also found that if the data is not sufficiently diverse, memorization (rather than any consistent generalization) is the preferred strategy.
+
+### Strengths
+* The paper adds to a rich literature on directly evaluating the inductive biases of language models. This paper approaches it from a grokking perspective, which, to my knowledge, has not been done before.
+* In Fig. 4, it is interesting that such a small variety of syntactic structures in the declaratives would lead to hierarchical generalization. I had assumed that a more naturalistic and varied distribution would be necessary to prevent memorization.
+* Thorough analysis of variation across many random seeds.
+
+### Weaknesses
+1. The findings could be much better contextualized with related work throughout the paper. For example, at L150: this is not the first work to analyze the effect of data on syntactic generalization. Mueller et al. (2023) (who are currently cited only in the Appendix) do a similar analysis, finding that simpler corpora lead to syntactic generalization with less data. Additionally, when referring to the syntactic transformations task setup at L144, Frank & Mathis (2007) [1] is the correct citation. At L154-161 (and App. F): it would be nice to contextualize this with analyses from McCoy et al. (2020) on instability across random seeds. In Sec. 4.3, these findings are very related to those of Papadimitriou & Jurafsky (2023) [2] (who are not cited) on training with recursive and/or cross-serial dependencies.
+2. The findings are somewhat obvious in light of the above citations, and in light of grokking work (which, by the way, should also be better cited and discussed; see [3,4,5]). Data determines generalization because neural networks are a statistical approach to learning. There are probably new insights to be gleaned from this new task setting compared to past grokking work, and it could be nice to explicitly enumerate these in the paper.
+3. L212-215: It is not clear why these hyperparameters were chosen, and why different hyperparameters were used for the two tasks.
+4. Related to the first point (but more minor), the decision of what to include in the main paper vs. the appendix Related Work sections currently feels arbitrary.
+
+### Questions
+Questions
+===
+1. Where did the hyperparameters in L212-215 come from? Was there some tuning involved, and if so, would it be possible to show results for other tested hyperparameters?
+2. By syntactic complexity, do you generally mean syntactic tree depth? The rarity/difficulty for humans in processing of particular structures? Number of edges in the tree (which could correlate with sentence length, even if the maximum depth of the tree were low)? I assume the first given the analyses in Sec. 4, but would be nice to explicitly define the term.
+3. L420: which studies?
+4. At first glance, Fig. 4 and Sec. 6 feel contradictory. There is some interesting nuance here that could be interesting to explore: diversity is necessary to prevent memorization, but at the same time, including a diverse mixture of complex and simple sentences leads to *worse* generalization than simply using more complex sentences. Why is this?
+
+Suggestions/Typos
+===
+* L182: there are *at least* two strategies. The model could rely on other heuristics, such as always moving the affirmative verb in QF as opposed to the main or first verb (as attested in Mueller et al., 2024).
+* Table 1 caption: by “left” and “right”, do you mean “top” and “bottom”?
+* L113: two periods
+
+References
+===
+[1] Robert Frank & Donald Mathis (2007). “Transformational networks.” Models of Human Language Acquisition. https://bpb-us-e2.wpmucdn.com/websites.umass.edu/dist/a/27637/files/2017/06/cogsci-2007.pdf
+
+[2] Isabel Papadimitriou & Dan Jurafsky (2023). “Injecting structural hints: Using language models to study inductive biases in language learning.” Findings of EMNLP. https://aclanthology.org/2023.findings-emnlp.563/
+
+[3] Yifei Huang et al. (2024). “Unified View of Grokking, Double Descent and Emergent Abilities: A Perspective from Circuits Competition.” COLM. https://arxiv.org/abs/2402.15175
+
+[4] Vikrant Varma et al. (2023). “Explaining grokking through circuit efficiency.” https://arxiv.org/abs/2309.02390
+
+[5] Ziming Liu, Eric J. Michaud, & Max Tegmark (2023). “Omnigrok: Grokking Beyond Algorithmic Data.” ICLR. https://arxiv.org/abs/2210.01117
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 2
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4
+
+### Summary
+This paper investigates the influence of training data composition on the generalization behaviors of language models, focusing on their ability to learn hierarchical syntactic representations versus relying on surface-level heuristics such as n-gram models. Through case studies involving English grammar tasks (specifically question formation and tense inflection) the authors show that the complexity and diversity of training data play pivotal roles in determining if models adopt hierarchical rules or simpler linear heuristics. 
+Interesting findings of the paper: 
+i) Adding sentences with deep syntactic trees in the training data encourages models to develop hierarchical syntactic representations  (and this seems to enable out of distribution generalization).
+ii) When the training dataset have a mix of simple and complex grammatical structures, models show unstable training dynamics and inconsistent rule commitments across different random seeds (this aligns with findings by McCoy et al. (2018, 2020))
+iii) In case of low data diversity, models tend to memorize patterns without learning robust hierarchical or linear rules, resulting in poor generalization.
+
+I think the main message of this paper is the relevance of training data features in shaping the inductive biases of neural networks.
+
+### Strengths
+- Interesting analysis of the mechanisms of rule commitment.
+- The identification of a memorization regime, where models stabilize without learning either hierarchical or linear rules.
+- The findings are replicated across two distinct grammatical tasks, and backed up by linguistics theroies.
+
+### Weaknesses
+- The experiments utilize relatively small transformer models trained on synthetic datasets with 100K samples (it doesn't make the study less valid, but larger models may exhibit different inductive biases and learning dynamics that are not captured by smaller-scale experiments).
+- The authors use a fixed set of hyperparameters (learning rate of 1e-4, Adam optimizer, specific layer configurations) across all experiments, without addressing how variations in hyperparameters might influence the model's ability to learn hierarchical rules or affect training stability.
+
+### Questions
+- It might be worthy study how other objectives (like MLM) interact with data composition to affect rule learning. 
+- Have you considered how the role of data composition might vary with languages that have different syntactic structures?
+
+### Soundness
+4
+
+### Presentation
+3
+
+### Contribution
+4
+
+---
+
+## Human Reviewer 3
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+3
+
+### Summary
+This paper investigate generalization of LMs.  Specilically, data composition significantly influences a model's out-of-distribution (OOD) generalization.  Insufficient data diversity can lead models to rely on memorization rather than achieving true generalization. They primarily validate these observations through learning on grammatical tasks.
+
+model: 12M  decoder-only Transformer.
+tasks: question formation, tense inflection.
+
+### Strengths
+The paper is well-written, with comprehensive visuals and thorough experimental results.
+
+### Weaknesses
+1. The conclusion in this paper have many overlaps with previous studies[1],[2],[3]; please specifiy your difference and contribution detailed.
+2. The experiments only contain two dataset: question formation and tense inflection. This limited validation is insufficient to support the overall conclusions of the paper, generally speaking.  More tasks and models are needed.  Suggest tasks: language modeling, mathematics reasoning … …
+3. The authors present only the experimental observation that "lower diversity data tends to promote memorization, while higher diversity data encourages generalization." These are  obvious conclusions compared with previous papers. Are there any new insights offering deeper theoretical reasoning or more controlled experiments to support these findings? or fuerther verfiication on LLMs, e.g. 7B ?
+
+### Grammar
+
+Line 113: “. . “
+
+Line 119: multiple abbreviation definitions
+
+Table 1: revise “Left” and “right” to “Top” and “Bottom”
+
+[1] Liu Z, Michaud E J, Tegmark M. Omnigrok: Grokking beyond algorithmic data[C]//The Eleventh International Conference on Learning Representations. 2022.
+
+[2] Zhu X, Fu Y, Zhou B, et al. Critical data size of language models from a grokking perspective[J]. arXiv preprint arXiv:2401.10463, 2024.
+
+[3] Wang B, Yue X, Su Y, et al. Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization[J]. arXiv preprint arXiv:2405.15071, 2024.
+
+### Questions
+See in Weaknesses.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+1

human_reviews/qL6brrBDk2.md

@@ -0,0 +1,188 @@
+# SAFLEX: Self-Adaptive Augmentation via Feature Label Extrapolation
+
+- Decision: Accept
+- Scores: 8, 5, 8, 8
+
+## Abstract
+Data augmentation, a cornerstone technique in deep learning, is crucial in enhancing model performance, especially with scarce labeled data.
+While traditional methods, such as hand-crafted augmentations, are effective but limited in scope, modern, adaptable techniques often come at the cost of computational complexity and are hard to fit into existing processes.
+In this work, we unveil an efficient approach that universally enhances existing data augmentation techniques by enabling their adaptation and refinement, thereby providing a significant and comprehensive improvement across all existing methods.
+We present \textbf{\SAFLEX} (\textbf{S}elf-Adaptive \textbf{A}ugmentation via \textbf{F}eature \textbf{L}abel \textbf{EX}trapolation), an approach that utilizes an efficient bilevel optimization to learn the \textit{sample weights} and \textit{soft labels} of augmented samples. This is applicable to augmentations from any source, seamlessly integrating with existing upstream augmentation pipelines.
+Remarkably, \SAFLEX effectively reduces the noise and label errors of the upstream augmentation pipeline with a marginal computational cost.
+As a versatile module, \SAFLEX excels across diverse datasets, including natural, medical images, and tabular data, showcasing its prowess in few-shot learning and out-of-distribution generalization.
+\SAFLEX seamlessly integrates with common augmentation strategies like RandAug and CutMix, as well as augmentations from large pre-trained generative models like stable diffusion. It is also compatible with contrastive learning frameworks, such as fine-tuning CLIP.
+Our findings highlight the potential to adapt existing augmentation pipelines for new data types and tasks, signaling a move towards more adaptable and resilient training frameworks.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This paper presents a principled method for data augmentation. To this end, the paper presents a bilevel optimization framework for weighing and soft-labelling the augmented data in order to compensate for the adverse generalization effects of weak, strong and sometimes meaningless augmented examples. Although the impact of data augmentation for generalization, in particular deep learning frameworks, has been substantial, there is still a lack of principled ways of doing data augmentation. This paper has identified this gap and convincingly addressed the problem.
+
+### Strengths
+The paper is well-written and easy to understand. 
+
+The diagrams and the equations are easy to follow.
+
+The experiments are performed on diverse datasets with various tasks, including medical imaging and tabular data.
+
+The results are highly encouraging.
+
+### Weaknesses
+A few important previous works on sampling and purifying GAN synthetic data are relevant to this paper.  It is important to acknowledge and discuss their contributions in the paper. 
+
+Caramalau, Razvan, Binod Bhattarai, and Tae-Kyun Kim. "Sequential graph convolutional network for active learning." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021.
+Bhattarai, Binod, et al. "Sampling strategies for gan synthetic data." ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020.
+
+### Questions
+I like the paper. Please see a few comments above.
+
+### Soundness
+4 excellent
+
+### Presentation
+4 excellent
+
+### Contribution
+4 excellent
+
+---
+
+## Human Reviewer 2
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+The paper argues that data augmentations can suffer with two main issues - 1. The augmented samples can become out of distribution to the training distribution and 2) the augmented samples can belong to a different class than the original sample. To tackle the first issue, the authors propose to add sample weights (w_i) to the augmented samples. Samples which are farther from the training distribution can be assigned a smaller weight. To tackle the second issue, the authors propose to make the one-hot label as soft-label to capture the uncertainties. 
+To learn the sample weights and the soft-label the authors pose a bi-level optimization problem where in the inner loop, the model parameters are optimized over the training and augmented samples and in the outer loop the optimal augmentation parameters are optimized for. 
+
+The authors conduct experiments across three settings - 1. medical datasets, 2. tabular datasets and 3. for contrastive learning approaches. Across all the experiments the authors show improved performance on top of standard augmentations such as RandAug, Mixup and CutMix.
+
+### Strengths
+1. The motivation in the paper about identifying the two issues with standard augmentation and then solving it by learning sample weights and soft-labels is really clear.
+
+### Weaknesses
+1. The main issue is a lack of proper baselines. Papers such as [1] have already explored using soft labels for augmentations where the softness is derived on the basis of augmentation strength. This paper's novelty thus gets limited. There is no comparison with [1] in any of the experiments. The authors should do a proper comparison with [1] and justify how their approach is better than it. 
+
+2. To solidify the experimental results the authors should also experiment with stronger architectures and datasets such as ResNet-101 over ImageNet as done in [1].
+
+I am willing to update my ratings if my concerns are addressed. 
+
+References - 
+
+1. Soft Augmentation for Image Classification. Liu et al. https://arxiv.org/pdf/2211.04625.pdf
+
+### Questions
+I have already mentioned it in the weakness section
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+2 fair
+
+---
+
+## Human Reviewer 3
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This article contributes a workflow named as SAFLEX as data augmentation. Here is a summry:
+
+(1) Authors unveil a novel parametrization for learnable augmentation complemented by an adept bilevel
+algorithm primed for online optimization.
+(2) Author's SAFLEX method is distinguished by its universal compatibility, allowing it to be effortlessly
+incorporated into a plethora of supervised learning processes and to collaborate seamlessly with an
+extensive array of upstream augmentation procedures.
+(3) The potency of authors' approach is corroborated by empirical tests on a diverse spectrum of datasets
+and tasks, all underscoring SAFLEX’s efficiency and versatility, boosting performance by1.2% on
+average over all experiments.
+
+### Strengths
+They have considered experiments of different data types and model training as downstream tasks, which demonstrate their workflow as a robust one.
+
+### Weaknesses
+From a model perspective, this is a good one as topic of adaptive learning, though a  little bit off the topic of this conference. 
+From data augmentation perspective, it is better to demo some more experiments in downstream task involves with high dimensional data.
+
+### Questions
+Is there any empirical experiments that SAFLEX can contribute to some other applicable downstream task like model training?
+
+### Soundness
+3 good
+
+### Presentation
+4 excellent
+
+### Contribution
+4 excellent
+
+---
+
+## Human Reviewer 4
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper addresses the training of high-performance neural networks on small amounts of training data by optimizing data augmentation. Existing methods in this field optimize the transformation itself in the feature space, which limits the available data augmentation transformations and has high computational complexity. This paper differs from these existing approaches by optimizing the importance weights of the input features and the soft labels to be assigned to the augmented data, thereby meta-learning the data augmentation available in the existing data augmentation pipeline. Focusing on features and labels in data augmentation has not received much attention, and the idea is novel. Furthermore, the paper proposes a method to approximate the bi-level optimization of meta-learning with validation data to a single-level optimization by borrowing the idea of gradient matching and developing an efficient algorithm. Experiments verify the effectiveness of the proposed method on various datasets, tasks, and combinations of data augmentation methods. On the other hand, the paper does not provide an evaluation of the first-order approximation or the computation cost, and there is room for improvement in this aspect.
+
+### Strengths
++ The paper proposes a novel data augmentation optimization strategy that optimizes the importance and labels of input features. The idea is very interesting and original.
++ The paper proposes a first-order approximation method to efficiently compute meta-learning that requires bi-level optimization.
++ The paper applies and evaluates the proposed method not only on image datasets but also on table datasets. This evaluation is important in supporting the paper's claim that the method can be applied to any data augmentation pipeline.
++ The paper provides experimental results on the recently widely used CLIP pre-trained models, effectively demonstrating the impact of the proposed method.
+
+### Weaknesses
+- Even though the paper proposes a first-order approximation method, it does not provide an evaluation of this method. In other words, the paper should provide a performance comparison with the usual bi-level optimization and a computation cost comparison with other data augmentation strategies such as Fast AutoAugment.
+- The writing of the paper is not necessarily of high quality. For example, Theorem 1 is very difficult to read because it contains multiple claims that make up the entire solution. Theorems and corollaries should be split for each claim, or if the propositions are ambiguous, they should be replaced with detailed explanations for each component, rather than in the form of a theorem. In fact, the proofs provided by the Appendix are almost obvious and make little theoretical contribution.
+
+### Questions
+- Do you think the proposed method can be applied to consistency-based semi-supervised learning with data augmentation, e.g., FixMatch [a]? The study of estimating the importance of samples and labels is well studied in the field of semi-supervised learning rather than in the field of data augmentation (e.g., FreeMatch [b]). If it can be shown that the scheme of the proposed method can be implemented in semi-supervised learning, the impact of this paper on the community will be even greater.
+
+[a] Sohn, Kihyuk, et al. "Fixmatch: Simplifying semi-supervised learning with consistency and confidence." NeurIPS 2020.
+
+[b] Wang, Yidong, et al. "Freematch: Self-adaptive thresholding for semi-supervised learning." ICLR 2023.
+
+### Soundness
+3 good
+
+### Presentation
+2 fair
+
+### Contribution
+3 good

human_reviews/qnWtw3l0jb.md

@@ -0,0 +1,179 @@
+# Fast Imitation via Behavior Foundation Models
+
+- Decision: Accept
+- Scores: 8, 8, 6, 8
+
+## Abstract
+Imitation learning (IL) aims at producing agents that can imitate any behavior given a few expert demonstrations. Yet existing approaches require many demonstrations and/or running (online or offline) reinforcement learning (RL) algorithms for each new imitation task. Here we show that recent RL foundation models based on successor measures can imitate any expert behavior almost instantly with just a few demonstrations and no need for RL or fine-tuning, while accommodating several IL principles (behavioral cloning, feature matching, reward-based, and goal-based reductions). In our experiments, imitation via RL foundation models matches, and often surpasses, the performance of SOTA offline IL algorithms, and produces imitation policies from new demonstrations within seconds instead of hours.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This paper leverages advances in behavior foundation models (BFMs) based on successor measures to build a BFM that can be pre-trained on a dataset of unsupervised trajectories without any prior knowledge of subsequent imitation behaviors. To that end, they describe several imitation learning methods that are based on a pre-trained Forward-Backward (FB) model. In the experiments the authors verify if an FB model that is pre-trained on one environment is able to imitate a wide range of tasks given only a few additional demonstrations on the target task.
+
+### Strengths
+1) The paper is easy to follow and well-structured with the main contributions listed clearly in the introduction alongwith grounding in related works and building blocks that make up this method. The paper is also technically sound with no methodical kinks that I could find. 
+2) Experimental details are sufficiently described for reproducibility.
+3) Experimental evaluation in the paper is very extensive covering a variety of methods with different pre-training methodologies on multiple tasks in the DeepMind Control suite.
+
+### Weaknesses
+Clarity: I think the clarity of the preliminaries can be improved. The paper assumes understanding of the FB framework which can make it hard to understand the core foundation of this work without going through previous papers. Effort should be made to make this work as self-contained as possible.
+
+### Questions
+Re future directions: Do the authors have any comments on utilizing FB pre-training for domain adaptation from sim to real? For example adapting to a locomotion task on a real robot from sim pre-training to overcome a low-dim parameter shift, and if this formulation will remain viable at all.
+
+### Soundness
+4 excellent
+
+### Presentation
+3 good
+
+### Contribution
+4 excellent
+
+---
+
+## Human Reviewer 2
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.
+
+### Summary
+This paper leverages recent work on the successor measure, i.e., a generalization of the successor representation to continuous (or large) state spaces. In particular, the authors leverage the Forward-Backward (FB) representation, which learns a low-rank decomposition of the successor measure conditioned on a policy embedding z. These policy embeddings should be approximately optimal for a particular reward function. The authors show how to leverage the FB representation to perform various forms of imitation learning by reusing the FB representation and learning an appropriate policy embedding for the imitation task.
+
+The work is well-motivated, and the importance is perhaps understated; imitation learning is a staple in many communities, and this paper provides notable improvements over well-established methods. As I see it, the primary contributions are:
+* The authors provide a novel application of the FB factorization of the successor measure to imitation learning.
+* Provide an extensive empirical analysis of FB applied to different forms of imitation learning, i.e., behavior cloning, reward-based imitation learning, distribution and feature matching, and goal-based imitation.
+
+### Strengths
+* The paper is very well written and easy to follow; each subsection detailing the application of FB is succinct but still provides many insights on the FB representation that prior work hadn’t touched on.
+* FB is unique in that it can be pre-trained from offline non-expert transitions and subsequently used to derive a reusable imitation policy from expert trajectories with very little overhead.
+* Well-designed empirical methodology with abundant baselines (both published and unpublished) on various continuous control domains. The experimental results are convincing and show significant improvement over the baselines.
+
+I also want to mention that I appreciate the level of detail in the supplementary materials and the lengths the authors went to describe how the baselines were implemented and tuned. This gives me confidence in the results and the ability to reproduce said results.
+
+### Weaknesses
+* I appreciate 21 tasks across four domains, but one concern is that the general findings might not extrapolate to other continuous control domains.
+* The results are limited to lower dimensional state spaces. I would appreciate at least discussing the method's limitations when scaling to higher-dimensional state spaces. It seems like FB has never been scaled beyond these low-dimensional continuous control tasks.
+
+### Questions
+* The DMC Maze tasks seem relatively straightforward; can you help give a better intuition as to why some methods fail to learn in this environment?
+
+### Soundness
+4 excellent
+
+### Presentation
+4 excellent
+
+### Contribution
+4 excellent
+
+---
+
+## Human Reviewer 3
+
+### Rating
+6
+
+### Rating Number
+6
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+This work leverages recent behavior foundation models (BFMs) that are based on successor features for fast imitation without RL or fine-tuning. Notably, this work uses the forward-backward (FB) framework to build BFMs that can solve any imitation learning tasks. The resulting set of algorithms is called FB-IL.  FB-IL accommodates several IL principles, including behavior cloning, feature matching, reward-based, and goal-based reductions. Experiments show that FB-IL can produce imitation policies from new demonstrations within seconds.
+
+### Strengths
+It is interesting and novel to leverage behavior foundation models based on successor features to enable fast imitation learning. Building on top of the forward-backward (FB) framework, FB-IL has several important and useful properties:
+
+1.	Pre-training the BFM only requires a dataset of unsupervised transitions/trajectories.
+
+2.	The BFMs solve imitation tasks without any complex reinforcement learning problem, and the computation is minimal,
+
+3.	The BFM is compatible with different imitation learning settings, which is impressive. 
+
+4.	The proposed FB-IL algorithms perform three orders of magnitude faster than offline IL methods, because FB-IL does not need to run full RL routines to compute imitation policies. 
+
+5. It is nice to see the proposed framework can be extended to IL tasks with non-stationary demonstrations.
+
+### Weaknesses
+1.	The proposed method is motivated and based on the previous work in zero-shot reinforcement learning (Touati et al., 2023). However, the forward-backward framework contents in section 3 are not easy to follow. Don’t assume readers are familiar with the framework proposed by Touati et al.
+
+2. The proposed methods are only evaluated in the tasks with low-dimensional states. I am wondering if the proposed methods still do not require RL routine or fine-tuning in environments with high-dimensional states.
+
+### Questions
+1.	This work leveraged BFMs based on successor measures, notably the forward-backward (FB) framework to solve fast imitation tasks. However, the introduction of successor measures is not clear enough. From section 3, we know that the successor measure describes the cumulated discounted time spent at each state following policy. The FB framework learns a tractable representation of successor measures that provides approximate optimal policies for any reward. What is the merit of achieving this goal?  And how does this property relate to FB-IL?
+
+2. Will the proposed methods require RL routine or fine-tuning in environments with high-dimensional states?
+
+### Soundness
+3 good
+
+### Presentation
+3 good
+
+### Contribution
+3 good
+
+---
+
+## Human Reviewer 4
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.
+
+### Summary
+The study extends prior research research into the Forward-Backward (FB) representation framework, a framework for training behaviour foundational models that hinges on representing agents + environment using an embedding space of policy representations and a set of two functions (F and B) that can be used to generate policy representations that are optimal with respect to emperically measuires rewards or reward functions unknown at model learning time.
+
+In this study, the framework is extended to address Imitation Learning (IL), where policies need to be learned from one or more expert demonstrations. The paper introduces multiple FB-IL methods, each mimicking existing IL approaches. However, these FB-IL models, are trained using pre-trained domain-specific FB models (that were not exposed to reward or demonstrations of tartidular tasks), reducing computation time significantly. This is because he model of the environment and how the agent can interact with it are pre-trained, and IL training primarily focuses on finding the right policy embedding (z).
+
+### Strengths
+The various proposed methods are extremely sample efficient (doing about as well with a single expert example as they do with 200), are orders of magnitude faster to compute (excluding the FB training time which is performed once per domain and can then be used learn any task within that domain, so long as the space was adequarely explored at FB training time).
+The results show that the IL-trained models can nearly match the performance of the expert models which they need to imitate (which were fully trained with RL).
+
+There are many IL methods and this paper aims to show that the FB framework can support them all. As such, the paper lists the main IL approaches and for each, introduces an FB-based variant that mimics that approach. In that respect this paper presents not one but **multiple** methods that are quite distict from one another yet all sare the same (FB) representation framework, thus driving home the point that the FB framework is quite versatile and does not limit the IL approaches.
+
+### Weaknesses
+The basic FB framework is skimmed though very quickly, requiring the unfamiliar reader to go back to 1-2 previous papers on the topic to fully inderstand the representational framework.
+
+The sheer number of "new" methods is a bit more that the reader might have bargined for in a 9-page paper (the very lengthy appendix serves as emphasis). Note that this can be seen as a strength of the work it self.
+
+Prior work on the FB framework has been limited to the somewhat simplistic gym environments, This work is no different. While the FB framework can scale by increasing the complexity of the embedding functions and the dimensionality of the policy representation space, it is unclear, at this point, that the framework can serve well in non-simulated environments. IL seems to me to be a practical method for real world problems and it would be nice to see FB applied to non-simulation (or more complex) environments.
+
+### Questions
+I found it quite surprizing that the method cannot improve generated policies in any significant way with more than a single IL example. I wonder if that is a reflection of the simplicity of the environments (given that it *is* doing about as well as the expert that generated the single example)?
+
+It was not clear to me how $\pi_𝑧(𝑠)= \arg \max _a  F (s, a, z)^T z$ is implemented. Is the arg max optimization occuring at every step or is the policy represented in yet another model that needs to be trained to estimate this arg max? If so how fast is that trainig?
+
+### Soundness
+4 excellent
+
+### Presentation
+3 good
+
+### Contribution
+4 excellent

human_reviews/r2nwBwodth.md

@@ -0,0 +1,189 @@
+# PFML: Self-Supervised Learning of Time-Series Data Without Representation Collapse
+
+- Decision: Reject
+- Scores: 3, 3, 3, 3
+
+## Abstract
+Self-supervised learning (SSL) is a data-driven learning approach that utilizes the innate structure of the data to guide the learning process. In contrast to supervised learning, which depends on external labels, SSL utilizes the inherent characteristics of the data to produce its own supervisory signal. However, one frequent issue with SSL methods is representation collapse, where the model outputs a constant input-invariant feature representation. This issue hinders the potential application of SSL methods to new data modalities, as trying to avoid representation collapse wastes researchers' time and effort. This paper introduces a novel SSL algorithm for time-series data called Prediction of Functionals from Masked Latents (PFML). Instead of predicting masked input signals or their latent representations directly, PFML operates by predicting statistical functionals of the input signal corresponding to masked embeddings, given a sequence of unmasked embeddings. The algorithm is designed to avoid representation collapse, rendering it straightforwardly applicable to different time-series data domains, such as novel sensor modalities in clinical data. We demonstrate the effectiveness of PFML through complex, real-life classification tasks across three different data modalities: infant posture and movement classification from multi-sensor inertial measurement unit data, emotion recognition from speech data, and sleep stage classification from EEG data. The results show that PFML is superior to a conceptually similar pre-existing SSL method and competitive against the current state-of-the-art SSL method, while also being conceptually simpler and without suffering from representation collapse.
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+This paper introduces a novel framework called Prediction of Functionals from Masked Latents (PFML) for modeling time-series
+data in self-supervised learning (SSL) pre-training. PFML stands out by eliminating the need for careful hyperparameter selection
+and mitigating representation collapse. Experimental results demonstrate that PFML outperforms a comparable SSL method and
+achieves competitive performance against SOTA SSL approaches.
+
+### Strengths
+1. The PFML method proposed in this paper effectively eliminates the cumbersome process of selecting hyperparameters, allowing the same model to be applied across different domains without the need for adjustments.
+2. The PFML model successfully avoids representation collapse during SSL pre-training, preventing the model from outputting a constant, time invariant representation. The paper provides detailed proof to support
+this claim.
+3. In the experimental section, the paper offers a thorough analysis of hyperparameters, the number of statistical functions, and the computational resources required for the proposed method.
+
+### Weaknesses
+1. The model proposed in this paper is an adaptation of the MAE model. The masking mechanism draws inspiration from wav2vec and data2vec, aiming to reconstruct the statistical information of the input signal. However, these contributions seem limited in scope.
+2. The experimental results of this model are only comparable to data2vec, which was published two years ago, suggesting that its performance may not surpass more recent methods in the field.
+
+### Questions
+1.Why were experiments conducted only on CNN and Transformer models, when the paper mentions that PFML is
+not limited to any specific model and can be applied to multiple sequence encoding models?
+2.The experimental section could be expanded to include additional results showcasing the application of PFML to other models, thereby demonstrating the broader applicability and universality of the PFML framework.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 2
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+4
+
+### Summary
+The paper presents PFML (Prediction of Functionals from Masked Latents), a self-supervised learning (SSL) algorithm designed to handle time-series data without suffering from representation collapse. The approach focuses on predicting statistical functionals (e.g., mean, variance) from masked embeddings instead of reconstructing the input signal, demonstrating comparable performance over similar SSL methods like MAE and Data2Vec in classification tasks for infant movement, speech emotion, and EEG-based sleep stages.
+
+### Strengths
+- Novelty in SSL for Time-Series: PFML introduces an innovative approach by predicting statistical functionals, reducing the complexity of reconstructing time-series data, a significant improvement over existing SSL methods.
+- Effective Across Multiple Modalities: The method is shown to work well across various domains, including IMU, speech, and EEG data, demonstrating its flexibility and robustness.
+- Simplicity in Hyperparameter Tuning: Unlike many SSL algorithms that require complex tuning, PFML is designed with simplicity, reducing the trial-and-error process in hyperparameter optimization.
+- Avoids Representation Collapse: PFML addresses a critical challenge in SSL, representation collapse, proving its reliability in training without requiring extensive countermeasures.
+
+### Weaknesses
+- Limited Exploration of Functionals: While PFML utilizes a set of many functionals, the choice of these functionals is somewhat arbitrary, and there is limited discussion on how different sets of functionals might affect performance across various datasets. As a matter of fact, many of the functionals are permutation-invariant like mean, variance, skewness and kurtosis directly computed from input data, while those computed from autocorrelation function can also be easily matched with a carefully permuted input data. This raises concerns about the technical soundness of the method, and it sets may not preserve the temporal structure of the data, which is crucial for time-series data.
+- No Data Augmentation: Despite recognizing the potential benefits of data augmentation during pre-training, the authors did not implement it to save computational time, which might limit the generalizability of results.
+- Narrow Benchmark Comparisons: The paper primarily compares PFML with MAE and Data2Vec. It would benefit from a broader range of comparisons with more diverse SSL methods across different tasks.
+- Lack of Theoretical Depth in Certain Areas: The theoretical justification for using functionals is not fully explored in terms of how it scales with different kinds of temporal variations or non-stationary data, potentially limiting understanding of the full potential of PFML.
+
+### Questions
+1. Could you provide an ablation study on the choice of functionals to demonstrate the impact of different sets of functionals on the performance of PFML across various datasets? Through comparison, it seems that the improvement in performance is marginal, which raises concerns about the necessity of using a large set of functionals and the potential concerns regarding effectiveness.
+2. Could you also provide additional comparison studies into the effectiveness of the proposed method for stationary and non-stationary data, and preferably include some synthetic datasets as well to study the design choice of functionals for different characteristics of data such as seasonality, trend, auto correlation, etc.?
+3. You proposed that representation collapse is critical in SSL, yet the paper does not provide a detailed discussion on how PFML effectively mitigates this issue, as MAE seems to not suffer from representation collapse already. And although data2vec do suffer from this issue, it is not clear how PFML is better than data2vec in this regard. Could you provide more insights into this? How do MAE and data2vec behave differently when one is experiencing representation collapse and the other isn't? Perhaps some visualization will help.
+4. The current evaluation of the method seems to be limited, in both the number of datasets and baselines. Could you provide more comprehensive evaluations on a broader range of datasets and SSL methods to better understand the generalizability of PFML? Such as, how does PFML compare to other SSL methods like [TimeCLR](https://doi.org/10.1016/j.knosys.2022.108606), [Time2Vec](https://arxiv.org/abs/1907.05321) and [TF-C](https://proceedings.neurips.cc/paper_files/paper/2022/hash/194b8dac525581c346e30a2cebe9a369-Abstract-Conference.html)?
+
+### Soundness
+1
+
+### Presentation
+2
+
+### Contribution
+1
+
+---
+
+## Human Reviewer 3
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+5
+
+### Summary
+The paper proposes a new style of self-supervised learning (SSL) with the core goal of avoiding representation collapse. It introduces a somewhat novel proxy task, where the model predicts statistical functionals of the masked segments instead of simply predicting the masked segment values. This approach aims to strike a balance between task simplicity and complexity. Their main claim is that this conceptually simple method effectively addresses the issue of representation collapse for time-series data.
+
+### Strengths
+1. The paper’s motivation is highly practical and deserves more attention from the time-series machine learning community.
+2. The simplicity-first design intent should be well-regarded in the machine learning community, and the paper supports this approach effectively.
+
+### Weaknesses
+1. The introduction of the paper misses several important references in the field of self-supervised learning including the very important non-contrastive methods since that is the whole premise of their motivation -- avoiding the choice of negative and positive anchors for contrastive learning [1]. In fact, previous works [2] have successfully leveraged SimSiam for audio/speech applications, which is one of the benchmark datasets in this work as well.
+
+2. I am curious about how the authors construct the functionals for multivariate time-series inputs.
+
+3. Lines 255-258 are challenging to follow, as the notations are not adequately explained. For example, in \( m_l \), what does \( l \) represent?
+
+4. Several claims are made without experimental substantiation:
+   - **4.1.** The authors state that the number of hyperparameters is fewer than in SOTA methods, but there is no objective data to support this. Also, the success of this method seems highly dependent on the window size for the masked embedding and the overall stochasticity of the underlying dataset.
+   - **4.2.** Echoing some previous concerns, Assumption 1 (lines 236-237) is overly strong and cannot serve as the foundation of this method. It is unreasonable to expect the dataset to present sufficient variability/nonstationarity to extract meaningful functionals and prevent collapse. What if I want to model a direct current (DC)-like signal, which is a component of my overall dataset?
+   - **4.3.** Some theoretical grounding to support the claim of avoiding collapse would increase confidence in this method’s applicability to general time-series data.
+
+5. The authors only consider two baselines. To genuinely claim the superiority of their method, they should evaluate it against several prominent time-series baselines Ts2Vec [3], BYOL [4] etc. Additionally, since the claim pertains to SSL representation learning, incorporating more diverse tasks (forecasting, anomaly detection, periodic target detection etc.) and datasets would help readers assess the method's utility for their applications.
+
+
+[1] "Exploring simple siamese representation learning." In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021.
+
+[2] "Efficient stuttering event detection using siamese networks." In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023.
+
+[3] "Ts2vec: Towards universal representation of time series." In Proceedings of the AAAI Conference on Artificial Intelligence, 2022.
+
+[4] "Bootstrap your own latent-a new approach to self-supervised learning." Advances in neural information processing systems 33 (2020).
+
+### Questions
+See weaknesses above.
+
+### Soundness
+1
+
+### Presentation
+2
+
+### Contribution
+1
+
+---
+
+## Human Reviewer 4
+
+### Rating
+3
+
+### Rating Number
+3
+
+### Confidence
+3
+
+### Summary
+The "PFML" paper presents a self-supervised learning (SSL) framework designed to address common challenges in time-series data, particularly avoiding representation collapse. The proposed model, Prediction of Functionals from Masked Latents (PFML), predicts statistical functionals rather than directly reconstructing the masked input signals. This approach is tested across three time-series data modalities—infant movement classification, emotion recognition, and sleep stage classification—demonstrating competitive performance and enhanced robustness against representation collapse compared to other SSL methods.
+
+### Strengths
+1.	PFML introduces an approach to SSL by predicting statistical functionals, effectively avoiding representation collapse, which is a common limitation in existing methods.
+2.	The method requires minimal hyperparameter tuning, making it more accessible for new data domains and modalities. Additionally, PFML shows strong resistance to representation collapse, as demonstrated in multiple data contexts.
+3.	PFML performs well across different time-series modalities, indicating broad potential for application in various fields such as healthcare and speech analysis.
+
+### Weaknesses
+1.	The motivation for using functionals as a target, while somewhat innovative, should be detailed to clarify its advantages over other SSL strategies, such as contrastive learning or clustering, in various time-series data types.
+2.	While PFML shows competitive results against baseline SSL methods, a more extensive comparison with recent SSL techniques for time-series data would better illustrate its relative strengths and areas for improvement.
+3.	The theoretical foundation of PFML is relatively limited, with insufficient rigorous derivation to support why predicting statistical functionals prevents representation collapse more effectively than other SSL methods. A more formal analysis would strengthen the paper and clarify the principles driving PFML’s effectiveness.
+I will reconsider my assessment after checking the rebuttal.
+
+### Questions
+1.	Could the authors provide more insight into how the choice of statistical functionals impacts PFML's adaptability across diverse time-series tasks?
+2.	How does PFML handle extremely noisy or incomplete data segments during pre-training?
+3.	Could this approach potentially extend to other types of sequential data, such as video or more structured clinical data like ECG?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2

human_reviews/sNntRFmn72.md

@@ -0,0 +1,192 @@
+# Cavia: Camera-controllable Multi-view Video Diffusion with View-Integrated Attention
+
+- Decision: Reject
+- Scores: 5, 8, 5, 5
+
+## Abstract
+In recent years there have been remarkable breakthroughs in image-to-video generation.
+However, the 3D consistency and camera controllability of generated frames have remained unsolved. Recent studies have attempted to incorporate camera control into the generation process, but their results are often limited to simple trajectories or lack the ability to generate consistent videos from multiple distinct camera paths for the same scene. To address these limitations, we introduce \textbf{Cavia}, a novel framework for camera-controllable, multi-view video generation, capable of converting an input image into multiple spatiotemporally consistent videos. Our framework extends the spatial and temporal attention modules into view-integrated attention modules, improving both viewpoint and temporal consistency. This flexible design allows for joint training with diverse curated data sources, including scene-level static videos, object-level synthetic multi-view dynamic videos, and real-world monocular dynamic videos. To our best knowledge, Cavia is the first of its kind that allows the user to precisely specify camera motion while obtaining object motion. 
+Extensive experiments demonstrate that Cavia surpasses state-of-the-art methods in terms of geometric consistency and perceptual quality. Project Page: \url{https://ir1d.io/Cavia/}
+
+## Human Reviews
+
+## Human Reviewer 1
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+3
+
+### Summary
+This paper presents a novel framework, Cavia, designed for generating multi-view videos with precise camera control. The authors introduce view-integrated attention mechanisms that enhance 3D consistency across perspectives. Additionally, Cavia is trained on a curated blend of static, monocular dynamic, and multi-view dynamic videos. Cavia demonstrates superior performance over baseline methods in both monocular video generation and cross-view consistency.
+
+### Strengths
+1. Novel Framework for Camera-Controllable Video Generation: This paper introduces an innovative framework aimed at generating videos with precise camera control.
+
+2. Superior Performance over Baselines: The proposed method demonstrates an  improvement over baseline models across key performance metrics.
+
+3. Good Monocular Setting Results: Under a monocular setting, the experimental results exhibit substantial improvements in the generated videos.
+
+### Weaknesses
+1. Limited Distinction in Visualization Results (Figure 5): The visual examples provided in Figure 5 lack noticeable differences, which undermines the claimed improvements. This inconsistency between visual and metric-based results calls into question the interpretability and significance of the enhancements.
+
+2. Poor Presentation: (i) The method section is minimal, with only three formulars, all of which are borrowed from references. The lack of detailed descriptions of the pipeline leaves the approach unclear and less accessible for readers. (ii) The paper includes non-standardized expressions, such as "(B V F C H W)" in Line 222. Additionally, Section 4 is ambiguously presented, making it difficult to comprehend.
+
+### Questions
+1. Visualization vs. Metric-Based Improvement: In Figure 5, the differences among the results appear minimal. What factors contribute to the significant metric-based improvements under the 2-view setting despite the subtle visual changes?
+
+2. GPU Memory Usage of Cross-Frame Attention: Given the extended token length in the cross-frame attention module, does this component impose a significant demand on GPU memory?
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 2
+
+### Rating
+8
+
+### Rating Number
+8
+
+### Confidence
+4
+
+### Summary
+This paper introduces CAVIA, a framework that advances video generator by adding camera controllability and enabling 4D generation. Specially, it focuses on generating multi-view consistent videos, allowing precise control over camera movement while maintaining object and scene consistency across frames. The Plücker embedding is adopted to inject camera features to the video generator. The cross-frame attention is used to enhancing temporal consistency within each video clip, and the cross view attention is utilized to enhance the content consistency between across different video clips. 
+CAVIA is also novel in its usage of multiple data sources, including static and dynamic, multi-view datas. This methods tested on various datasets and compared with some existing methods, demonstrating its state-of-the-art performance.
+
+### Strengths
+1. Writing is good, and is easy to follow.
+2. The model is simple and effective, and easy to use the existing pre-trained video generation models.
+3. The data preparation process is logical and effective, it leverages a mix of statics multi-view, monocular videos, and synthetic multi-view datas. The model benefits from such a broader range of training data, enhancing its generalization and robustness.
+4. Extensive experiments demonstrate the effectiveness of each design choice.
+
+### Weaknesses
+The main weaknesses are in the experiment part:
+1. What are the "general" videos come from? Compared to the RealEstate10K dataset, does the "general" videos have larger camera motion? And does the dynamic complexity is limited?
+2. In Table 2, there should be some quantitative comparisons with the CVD model.
+3. In the provided videos, the object dynamic degree is limited, can you provide a comparison on the object dynamic degree between CAVIA and the base video generation model SVD?
+
+### Questions
+See weaknesses
+
+### Soundness
+4
+
+### Presentation
+3
+
+### Contribution
+3
+
+---
+
+## Human Reviewer 3
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+4
+
+### Summary
+This paper presents a novel method for camera-controllable multi-view video generation. Leveraging a pre-trained I2V model, the method introduces camera-conditioned Plücker coordinates in the latent space for conditional control. Additionally, it extends spatial attention into cross-view attention and promotes 1D temporal attention into cross-frame 3D attention. These enhancements have been evaluated and found beneficial for achieving cross-view and cross-frame consistency, as well as precise camera control.
+
+To address the challenge of limited multi-view video data, an effective joint training strategy is proposed, utilizing a curated mixture of static, monocular dynamic, and multi-view dynamic videos. Experiments demonstrate the superiority of the proposed method over existing competitors in terms of multi-view consistency and camera control precision. However, it is worth noting that most of the demonstrated examples pertain to specific domains, involving minimal dynamics and simple backgrounds, which raises questions about the method's performance in more general scenarios.
+
+### Strengths
+1. From the qualitative and quantitative results, it is evident that this method indeed offers superior camera trajectory control accuracy. This advantage is attributed to the benefits brought by 3D temporal attention (maybe also fulltuning) and the carefully collected training dataset.
+
+2. To overcome the scarcity of multi-view video dataset, a curated mixture of existing available static, monocular dynamic, and multi-view dynamic videos is exploited via joint training scheme. This provides a solution for addressing data issues in future research on this task.
+
+3. The experimental evaluation and comparison are comprehensive, evaluating the geometric accuracy of camera control from multiple perspectives. It demonstrates noticeable superiority of the proposed method.
+
+### Weaknesses
+1. The technical novelty is mild. Adopting plucker embedding for camera control is not new. Extending spatial attention for cross-view attention is also straightforward. Anyway, it is inspiring that 3D attention instead of 1D temporal attention is highly desired to help camera control accuracy, and this is well validated by multiple experiments.
+
+
+2. One of the major weakness is the less convincing results. 
+The most important aspects of multi-view video generation are motion synchronization and geometric consistency. However, most of the examples presented in the paper involve minimal object movement (either a single object with a simple, weakly moving background or entirely static scenes), and the camera angles do not vary significantly, with most shots taken from the same side of the object. This makes it difficult to evaluate the proposed method's capabilities in this task.
+
+For example, I would like to see examples like this: "on a street with vehicles and pedestrians, a horse is running, with one camera capturing a global view from behind the horse and another camera tracking from the side". Showing more examples like this would be more convincing and demonstrate the value of controllable multi-view video generation.
+
+
+3. In my understanding, making use of monocular videos could promote the model's generalization. But most of the cases in the paper, are close to the style of Objeverse dataset. Is it because the fulltuning destroyed the capabiliy of SVD in generating realistic style and scene? If so, this should be claimed as a limitation, especially compared to those adapter plugin methods like MotionCtrl and CameraCtrl.
+
+
+4. From the results in Fig. 8, it appears that while the introduction of monocular video increases generalization ability, such as maintaining the content of the input images, it also affects the accuracy of multi-view camera control. In the second row of (b), the shooting direction for the fish and the fox did not turn as expected.
+
+### Questions
+1. In Table 3, I'm curious why the impact of "w/o cross-view" is even smaller than that of "w/o cross-frame." Intuitively, cross-view should be an essential component for multi-view setups (without it, it's challenging to maintain multi-view consistency and motion synchronization), whereas cross-frame primarily enhances the accuracy. Does this result imply that most of the outcomes are actually static, and similar results could be achieved through independent monocular camera control?
+
+2. Is your model fully tuned? What is the parameter count?
+
+3. Are the results of the ablation study in Table 3 based on monocular camera control or multi-view camera control? It would be advisable to evaluate these two settings separately.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+---
+
+## Human Reviewer 4
+
+### Rating
+5
+
+### Rating Number
+5
+
+### Confidence
+5
+
+### Summary
+This paper introduces Cavia, a novel framework for generating multi-view videos with camera controllability. It incorporates view-integrated attention mechanisms—specifically, cross-view and cross-frame 3D attentions—to enhance consistency across different viewpoints and frames. The paper also presents an effective joint training strategy that leverages a curated mix of static, monocular dynamic, and multi-view dynamic videos. This approach ensures geometric consistency, high-quality object motion, and background preservation in the generated results. Experimental results show that Cavia outperforms baseline methods in terms of geometric and perceptual quality for monocular video generation and cross-video consistency. Additionally, the flexible framework can process four views at inference, leading to improved view consistency and enabling 3D reconstruction of the generated frames.
+
+### Strengths
+- The Cross-View Attention and Cross-Frame Attention mechanisms are intuitively designed and effectively enhance temporal consistency. The incorporation of 3D attention particularly strengthens this aspect.
+
+- The joint training strategy is commendable for its use of diverse dataset sources, including monocular videos and multi-view videos, which contributes to the robustness of the model.
+
+### Weaknesses
+- The video content presented demonstrates low dynamics (object), which may limit the ability to fully showcase the capabilities of the proposed method.
+
+- The contribution of this work appears to be incremental. The proposed method shares significant similarities with CVD, suggesting that the innovations may be more evolutionary than revolutionary.
+
+- The trajectories showcased in the supplementary demo are relatively simple, which raises questions about the model's ability to generalize to more complex scenarios. Further demonstration on a broader range of data would help to substantiate the method's effectiveness.
+
+### Questions
+- The camera conditions are normalized to a unit scale, can this approach still achieve variable speed camera control similar to MotionCtrl?
+
+- Is it sufficient for the filter method used in Particle-sfm for camera pose estimation to rely on the number of points for accurate pose estimation in videos with significant object motion?
+
+- It would be beneficial to include data samples from various sources to enhance understanding. Additionally, filtering out poor quality samples is necessary.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2