mesh, which is usually referred to as mesh reconstruction, rather than mesh generation. Though a transformer-based autoregressive architecture is applied, I still don't think it is appropriate to name the paper a mesh generation paper. \\n\\n- It is unfair to use meshes generated by Rodin as the shape conditioning for comparison with MeshGPT. From my understanding, the Rodin engine largely (or purely) provides the generation ability while MeshAnything only acts as a remesher component.\\n\\n- Another concern is that the number of AM mesh faces generated by MeshAnything is limited to a maximum of 800. Using such a limited number of faces makes it insufficient for representing meshes with more complex structures.\\n\\n- This work appears to have stringent data requirements, as the authors were only able to filter 56k high-quality artist-created meshes from the Objaverse and ShapeNet datasets for training. So it may be challenging to scale this method up to more data and modeling parameters (but it might be sufficient to train a point cloud reconstructor/mesh remesher).\",\n \"questions\": \"My suggestion is to change the description of mesh generation to mesh reconstruction/remeshing. \\n\\nI regard the proposed method as a novel learning-based remesher (AR applied), but I don't think the paper could be classified into a generation paper.\"\n },\n {\n \"id\": \"05N6vnZse4\",\n \"initial_rating\": 6,\n \"confidence\": 5,\n \"soundness\": 3,\n \"contribution\": 3,\n \"presentation\": 3,\n \"summary\": \"This paper tackles the important question of converting various 3D representations to explicit mesh representations similar to what artists create. For this, the authors adopt the idea of MeshGPT and inject shape-conditioned information to make the model focus on learning topology. To make the model agnostic to specific 3D representations, e.g., NeRF or Gaussian Splats, the authors choose to convert all representations to point clouds and generate the mesh based on the features from the point cloud. Further, the authors develop a fine-tuning strategy for the VQ-VAE's decoder to create a noise-resistant one to reduce the gap between the data in training and real usages. A user study demonstrates the effectiveness of the proposed approach.\",\n \"strengths\": \"- originality-wise: the choice of converting all representations to a point cloud and using the point cloud to generate corresponding mesh is interesting;\\n- quality-wise: qualitative results are of high quality;\\n- clarify-wise: the paper is well-written and easy to follow;\\n- significance: to produce a mesh similar to what artists create is of great importance to bridge the current 3D generation research into real-world usage.\",\n \"weaknesses\": \"## Concerns about the user study\\n\\nSince the goal of the project is to produce artist-created meshes, it is important to make sure that the model's output is aligned with real artists' preferences. However, there is no information about whether the 41 users who participated in the study (L517) are qualified especially Tab. 1 is the main results reported in the paper. For \\\"qualified\\\" users, I mean people who are artists with enough experience in the field of 3D creation or editing, etc.\\n\\nCan authors clarify? If the users are not qualified, I would not be convinced by the result.\\n\\n## Concerns about comparison with mesh generation pipeline\\n\\nIn L508, the authors state:\\n> MeshAnything is a shape-conditioned mesh generation method, we sampled shapes randomly from the evaluation set of Objaverse as inputs for MeshAnything, while for the baseline methods, we performed random sampling directly.\\n\\nFurther, in L862, the authors state:\\n> We quantitatively evaluate mesh quality by uniformly sampling 100K points from the faces of both the ground truth meshes and the predicted meshes, and then computing a set of metrics to assess various aspects of the reconstruction.\\n\\nIf I understand correctly:\\n- the procedure for evaluating MeshAnything is: 1) sample a mesh and then sample points from the evaluation set; 2) run MeshAnything on the sampled point cloud to obtain a predicted mesh; and 3) sample points from the predicted mesh.\\n- the procedure for the baselines are 1) directly run the generation; and 2) sample points from the generated mesh.\\n\\nSuch a setup is quite unfair in my opinion. Essentially, MeshAnything generates meshes aligned with the evaluation dataset. I think it will surely perform well in terms of those metrics for evaluating generation qualities in Tab. 2. Especially on Objaverse, which has tons of various categorical objects, the unfairness will be amplified.\\n\\nI also agree that it is not easy to compare with baselines as MeshAnything needs to have a point cloud conditioning. ShapeNet may be more suitable due to the constrained categories.\\n\\n## Questions about generation procedure\\n\\nCan the authors clarify how the generation is conducted? As a generative framework, should the model output various plausible meshes based on the input point cloud? However, in the paper, there is always one mesh corresponding to the input. Does this mean that the authors always use **max** logits during the sampling, i.e., greedy search? Can authors provide more generation variants, e.g., with beam search? I would like to know how diverse the mesh generation could be.\\n\\nFurther, a question related to the evaluation on Tab. 2 is whether the authors only use **max** sampling to evaluate. \\n\\n## Insufficient qualitative results\\n\\nThere are actually no qualitative comparisons to baselines, e.g., MeshGPT and PolyGen in the paper. Please provide. \\n\\n## Citation format\\n\\nPlease refer to the template instructions and use the correct citation commands. Currently, all citations are with `\\\\citet` instead of `\\\\citep`. They look so weird as those citations appearing without parenthesis break the sentences.\",\n \"questions\": \"See \\\"weakness\\\".\"\n },\n {\n \"id\": \"qyEsyD6A5F\",\n \"initial_rating\": 8,\n \"confidence\": 4,\n \"soundness\": 3,\n \"contribution\": 4,\n \"presentation\": 4,\n \"summary\": \"This paper presents MeshAnything, a model framing mesh extraction as a generative task, resulting in artist-crafted meshes that align with specified shapes. The approach first establishes a mesh vocabulary through VQ-VAE, followed by training a shape-conditioned, decoder-only transformer with a noise-resistant decoder design on the learned vocabulary, enabling shape-conditioned autoregressive mesh generation. This paper is well-crafted, and I anticipate that it will positively influence the field of 3D content generation though the pipeline appears somewhat straightforward. The performance of the method seems promising.\",\n \"strengths\": \"This is a well-written paper that clearly conveys the authors' contributions. The topic is highly relevant to the field, as it addresses the gap between generated shapes and practical applications - I expect the authors will consider open-sourcing their code to facilitate further research. The effectiveness of the method is demonstrated in the experimental sections.\",\n \"weaknesses\": \"I recommend that the authors showcase real-world application examples to highlight the significance of their research. For example, the authors could demonstrate how Artist-Created Meshes (AMs) improve shape manipulation for artist-driven shape modifications or offer enhanced rendering performance (e.g., efficiency) compared to conventional dense meshes.\\n\\n## Related Works\\nAdditionally, recent advancements in 3D reconstruction and generation could be acknowledged, such as:\\n- Part123: Part-aware 3D Reconstruction from a Single-view Image (SIGGRAPH 2024)\\n- SyncDreamer: Generating Multiview-consistent Images from a Single-view Image ICLR 2024.\\n- CLAY: A Controllable Large-scale Generative Model for Creating High-quality 3D Assets (SIGGRAPH 2024)\\n\\n\\n## Exposition\\n\\n- \\\"MeshAnything converts any 3D representation into Artist-Created Meshes (AMs), i.e., meshes created by human artists.\\\" Consider rephrasing \\\"any 3D representation\\\" to \\\"different 3D representations.\\\"\\n\\n- \\\"This is because all meshes produced by these methods (Lorensen & Cline, 1987; Chernyaev, 1995; Lorensen & Cline, 1998; Shen et al., 2021b; Chen et al., 2022; Shen et al., 2023) exhibit significantly poorer topology quality compared to AMs. As shown in Fig. 2, these methods rely on dense faces to reconstruct 3D shapes, completely ignoring geometric characteristics.\\\" The logic here seems unclear. Generally, mesh topology pertains to the connectivity of vertices, while geometric features relate to vertex positions. Could the authors clarify how \\\"poorer topology quality (relying on dense faces)\\\" affects \\\"geometric characteristics\\\"?\",\n \"questions\": \"1. Will the data and code be available for reproducibility?\\n2. Can the limitations and potential future work be expanded in the main paper? Are there any failure cases the method cannot handle? Additionally, what would happen if watertightness is required during generation? Including these discussions could enhance the paper.\\n3. \\\"Due to its generative nature, our method is not as stable as reconstruction-based mesh extraction methods.\\\" Could the authors clarify the meaning of \\\"stable\\\" in this context?\\n4. How does the method perform on open surfaces?\"\n }\n]","isConversational":false},"scores":{"kind":"string","value":"{\n \"rating\": \"5;5;6;8\",\n \"rating_avg\": 6,\n \"confidence\": \"5;3;5;4\",\n \"confidence_avg\": 4.25,\n \"soundness\": \"3;2;3;3\",\n \"soundness_avg\": 2.75,\n \"contribution\": \"3;2;3;4\",\n \"contribution_avg\": 3,\n \"presentation\": \"4;3;3;4\",\n \"presentation_avg\": 3.5\n}"},"metadata":{"kind":"string","value":"{\n \"primary_area\": \"\",\n \"track\": \"main\",\n \"venue\": \"ICLR 2025 Poster\",\n \"venueid\": \"ICLR.cc/2025/Conference\",\n \"processed_at\": \"2026-01-14T22:16:01.732885\"\n}"},"meta_review":{"kind":"string","value":"{\n \"id\": \"0Dfg3QoiWG\",\n \"metareview\": \"The submission received somewhat mixed reviews. The reviewers appreciate the work addressing artist-like shape generation for practical artistic applications, and strong results were presented. The major concerns from 7tHF and tUZi were on doubts on scalability and missing/unfair comparisons. The AC carefully read through the paper, the reviewers' comments, the authors' rebuttal and the discussions. While reservations for scalability remains, the contributions of the paper includes a new method that address a practical problem with strong preliminary results. The AC believes the merits outweighs the limitations in general; as such, the AC recommends acceptance.\",\n \"additional_comments\": \"The reviewers raised questions mostly regarding scalability (7tHF, tUZi), missing/unfair comparisons (7tHF, tUZi, qFPE, 3Tqd), and task framing of reconstruction vs. generation (tUZi). Concerns around scalability and comparisons have been partially addressed. On tUZi's question on reconstruction vs. generation, the AC agrees with the authors that the work can be considered a 3D generative model. The AC disregards any discussion on citing follow-up works during this process.\"\n}"},"decision":{"kind":"string","value":"{\n \"decision\": \"Accept (Poster)\"\n}"}}},{"rowIdx":1299,"cells":{"paper_id":{"kind":"string","value":"KHTkRhq2aB"},"arxiv_id":{"kind":"string","value":"2406.17923v1"},"title":{"kind":"string","value":"PAFT: A Parallel Training Paradigm for Effective LLM Fine-Tuning"},"markdown":{"kind":"string","value":"{\n \"content\": \"## Abstract\\n\\nAbstract Large language models (LLMs) have shown remarkable abilities in diverse natural language processing (NLP) tasks. The LLMs generally undergo supervised fine-tuning (SFT) followed by preference alignment to be usable in downstream applications. However, this sequential training pipeline leads to alignment tax that degrades the LLM performance. This paper introduces PAFT, a new PA rallel training paradigm for effective LLM F ine- T uning, which independently performs SFT and preference alignment (e.g., DPO and ORPO, etc.) with the same pre-trained model on respective datasets. The model produced by SFT and the model from preference alignment are then merged into a final model by parameter fusing for use in downstream applications. This work reveals important findings that preference alignment like DPO naturally results in a sparse model while SFT leads to a natural dense model which needs to be sparsified for effective model merging. This paper introduces an effective interference resolution which reduces the redundancy by sparsifying the delta parameters. The LLM resulted from the new training paradigm achieved Rank #1 on the HuggingFace Open LLM Leaderboard 1 1 1 https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard . Comprehensive evaluation shows the effectiveness of the parallel training paradigm.\\n\\n## 1 Introduction\\n\\nIn recent years, large language models (LLMs) have emerged as the standard approach to addressing natural language processing (NLP) tasks. The typical way of building an LLM for downstream applications generally follows a sequential training pipeline consisting of two phases: 1. Supervised Fine-tuning (SFT), where the pre-trained LLM is fine-tuned with the language modelling loss on demonstrations of the desired behaviour. 2. Alignment with human preference, where the model produced by the SFT phase is further fine-tuned with an alignment algorithm like Reinforcement Learning from Human Feedback (RLHF) or Direct Preference Optimization (DPO), etc. While this sequential pipeline has been used to seemingly great success, how the SFT and the preference alignment work better with each other is underexplored.\\n\\nRecent studies OpenAI (); Askell et al. (); Song et al. () have found that the preference alignment phase can cause the LLM to forget the diverse capabilities that it has acquired from earlier phases, despite aligning the LLM with human expectation. This phenomenon, also known as the *alignment tax* in the literature Ouyang et al. (), has accumulated substantial attention from both academia and industry. The alignment tax inherently results from catastrophic forgetting present in the staged training. To reduce catastrophic forgetting and thus alignment tax, this paper introduces a new parallel training paradigm for LLM fine-tuning, named PAFT, which independently performs SFT and preference alignment with the same pre-trained model on respective datasets, instead of sequentially conducting SFT followed by preference alignment. The model from SFT and the model from preference alignment are then merged into a final model by parameter fusing for use in downstream applications.\\n\\nAs discovered by prior work Yadav et al. (); Yu et al. (), direct model merging causes the parameter values to interfere across models, thereby harming the performance of the final model. The interference, which reduces parameter magnitudes in the merged model and eliminates subtle distinctions among values, can attribute to the redundant *delta parameters*, i.e., the differences in values between fine-tuned and pre-trained parameters, resulted from fine-tuning. Previous studies on model pruning Hoefler et al. (); Thimm and Fiesler () have shown that during fine-tuning, many model parameters can change over the course of fine-tuning but only have a small impact on performance. However, when merging a parameter that is influential for one model but redundant (i.e. not influential) for other models, the influential value may be obscured by the redundant values, lowering the overall model performance. This work reveals the dense properties of the delta parameters resulted from SFT. To mitigate the dense property of SFT, we propose an effective interference resolution which reduces the redundancy by sparsifying the delta parameters by adding a L1-norm penalty to the original SFT loss function. The existing findings indicate that the inclusion of the L1 term enhances the sparsity of the SFT. This method of implicitly inducing sparsity has been evaluated against a technique that introduces sparsity explicitly, i.e., DARE Yu et al. (), demonstrating the advantages of employing the L1-norm on LLM’s performances in downstream tasks.\\n\\nFinally, the sparse delta parameters from SFT and preference alignment are merged into a single stronger model. Different merging methods are assessed, and TIES and Task Arithmetic are shown to be the best model merging methods, depending on base models. The method of Parallel $\\\\text{SFT}_{\\\\text{sparse}}$+DPO merged through TIES based on Mistral-7B sets a new benchmark for 7B models, i.e., 0.6524 on average over the six tasks in HuggingFace Open LLM Leaderboard. Notably, Parallel $\\\\text{SFT}_{\\\\text{sparse}}$+DPO consistently outperforms Parallel SFT+DPO across all model merging methods, showing the effectiveness and robustness of the PAFT training paradigm.\\n\\nThe contributions of this paper are threefold:\\n\\n- 1.\\nEvidence is presented that parallel training of SFT and preference alignment outperforms sequential training, effectively reducing the alignment tax.\\n- 2.\\nThe significance of sparse model integration is highlighted as a mean to prevent model conflict while preserving the full capability of each model. We demonstrate the superiority of the L1-norm over DARE as a more effective and higher-quality method for promoting sparsity in model training across various model merging techniques.\\n- 3.\\nWe conduct comprehensive evaluation of PAFT on well-known public benchmarks including Open LLM Leaderboard and AlpacaEval. The PAFT-ed 7B model achieved Rank #1 in the 7B/8B model category on the Open LLM Leaderboard, and the PAFT-ed 70B model topped the Leaderboard globally.\\n\\nFigure: (a) Staged training\\nRefer to caption: extracted/5691539/figs/staged.png\\n\\n## 2 Methodology\\n\\n### 2.1 Problem Setting\\n\\nGiven a pre-trained LLM, such as Mistral and Llama, we aim to optimize the model for a wide range of downstream tasks by fine-tuning it either fully or with parameter-efficient tuning such as LoRA Hu et al. (), using SFT and preference alignment. Throughout this paper, $\\\\theta$ denotes the trainable parameters; $\\\\theta_{\\\\mathrm{pre}}$ denotes the parameters of the pre-trained model; $\\\\theta_{\\\\mathrm{sft}}$ denotes the parameters of the model fine-tuned with SFT; $\\\\theta_{\\\\mathrm{xpo}}$ denotes the parameters of the model fine-tuned with preference alignment, such as PPO Schulman et al. (); Ziegler et al. (), DPO Rafailov et al. () and ORPO Hong et al. (), etc.; $\\\\delta_{\\\\mathrm{sft}}=\\\\theta_{\\\\mathrm{sft}}-\\\\theta_{\\\\mathrm{pre}}$ denotes the delta parameters between the SFT-ed model and the pre-trained model; and $\\\\delta_{\\\\mathrm{xpo}}=\\\\theta_{\\\\mathrm{xpo}}-\\\\theta_{\\\\mathrm{pre}}$ denotes the delta parameters between the preference-aligned model and the pre-trained model.\\n\\n### 2.2 Parallel Training\\n\\nSFT and preference alignment are two distinct methodologies designed to enhance the capabilities of pre-trained LLMs for specific applications. SFT focuses on boosting the performance of LLMs on downstream tasks by fine-tuning them with datasets that closely resemble the target task. This process tailors the model’s responses to be more accurate and relevant for a specific use-case. In contrast, preference alignment, such as RLHF, DPO and ORPO, etc., is a methodology that refines a model’s outputs based on human preferences. It generally fine-tunes the model on pairs of responses to an input query, one of which is preferred over the other one. Preference alignment uses such feedback signal to guide the model towards generating outputs that align with human expectation and ethical standards. This approach is particularly valuable for addressing the ethical considerations that arise when deploying LLMs in real-world scenarios.\\n\\nNowadays, researchers have applied SFT to enhance the performance of LLMs on targeted tasks, and then employed preference alignment to further align the models with human preferences. However, this sequential application of SFT followed by preference alignment has often led to a compromise in task-specific performance - a phenomenon referred to as the alignment tax. This occurs because the distinct objectives of SFT and preference alignment can sometimes be at odds, with the alignment process potentially undoing some of the task-specific optimizations achieved through SFT.\\n\\nWe address the challenge of the alignment tax by a novel approach that involves SFT and preference alignment concurrently using adapter training, such as LoRA Hu et al. (). This method takes full advantages and strengths of both SFT and preference alignment without sacrificing performance in either one, i.e., ensuring that the resulting model maintains high performance in downstream tasks while also being aligned with human preferences, thus overcoming the limitations associated with the alignment tax. During the training process specifically, based on the same pre-trained model $\\\\theta_{\\\\mathrm{pre}}$, the two separate adapter parameters, denoted as $\\\\delta_{\\\\mathrm{sft}}$ and $\\\\delta_{\\\\mathrm{xpo}}$, are learned in parallel from downstream ground truth and human preferences, respectively. The proposed PAFT seeks to merge the $\\\\delta_{\\\\mathrm{sft}}$ and $\\\\delta_{\\\\mathrm{xpo}}$ in an effective way of avoiding feature interference. Figure 1 compares the typical staged training pipeline and our parallel training pipeline PAFT.\\n\\n### 2.3 Sparse Merging\\n\\nThe integration of dense neural network models often results in a suboptimal combined model due to the phenomenon of parameter interference. This challenge has led researchers to explore alternative strategies. Our investigations reveal that by increasing sparsity of a fine-tuned adapter, the performance of merging the adapter with the base model can be improved. Specifically, the parameter $\\\\delta_{\\\\mathrm{xpo}}$, derived from adapter training like LoRA, demonstrates clear sparsity, as depicted in Figure 2. In contrast, the sparsity in a SFT adapter, denoted by $\\\\delta_{\\\\mathrm{sft}}$, is not pronounced. To increase the sparsity within $\\\\delta_{\\\\mathrm{sft}}$, we propose the incorporation of an L1 regularization term during the SFT process. This modification to the fine-tuning procedure is expressed mathematically as follows:\\n\\n$$ $L_{\\\\text{$\\\\text{SFT}_{\\\\text{sparse}}$}}=L_{\\\\text{SFT}}+\\\\lambda\\\\cdot\\\\|\\\\delta_{ \\\\mathrm{sft}}\\\\|_{1}$ (1) $$\\n\\nHere, $L_{\\\\text{SFT}}$ represents the conventional cross-entropy loss function, and $\\\\lambda$ is a weighting factor that controls the strength of the sparsity regularization. Our results indicate that this approach significantly enhances the sparsity of $\\\\delta_{\\\\mathrm{sft}}$, with sparsity levels over 90%, as illustrated by the SFT_sparse in Figure 2.\\n\\nGiven sparse representations for adapters of both SFT and preference alignment, the challenge is to effectively merge these delta parameters, $\\\\delta_{\\\\mathrm{sft}}$ and $\\\\delta_{\\\\mathrm{xpo}}$, with the original pre-trained model, $\\\\theta_{\\\\mathrm{pre}}$, while preserving the performance benefits of SFT and preference alignment. The merging process can be formalized by the equation:\\n\\n$$ $\\\\theta_{\\\\mathrm{merge}}=f(\\\\theta_{\\\\mathrm{pre}},\\\\delta_{\\\\mathrm{dpo}},\\\\delta_{ \\\\mathrm{sft}})$ (2) $$\\n\\nIn our study, we explore a variety of merging methods proposed in the literature, including SLERP, Task Arithmetic, TIES, DARE TIES, and Linear. Detailed discussions of these merging methods are provided in the Related Work section.\\n\\nFigure: Figure 2: Adapter sparsity for SFT and DPO. The sparsity levels are computed by first merging the parameters from LoRA matrices $\\\\delta_{A}$ and $\\\\delta_{B}$ through matrix multiplication ($\\\\delta=\\\\delta_{B}\\\\times\\\\delta_{A}$), and computing the percentage of elements within $\\\\delta$ that are less than a threshold of $1\\\\times e^{-5}$, indicating the proportion of weights approaching zero. The reported sparsity is the average across all layers.\\nRefer to caption: extracted/5691539/figs/sparsity.png\\n\\n## 3 Experiments\\n\\n### 3.1 Evaluation Settings\\n\\nIn this study, we conduct comprehensive evaluation on both the Open LLM leaderboard provided by HuggingFace and the AlpacaEval benchmark. The Open LLM Leaderboard benchmark suite encompasses a diverse set of six benchmark tasks, namely ARC, HellaSwag, MMLU, TruthfulQA, Winogrande, and GSM8K, along with their aggregated performance metrics.\\n\\nIn our experiments, we employ two state-of-the-art pre-trained models: Mistral-7B Jiang et al. () and Llama-3-8B(^2^22Note that while the Llama 3 model is referenced in our work, the official documentation for this model has not been released at the time of writing, and thus we cite its official GitHub site as a proxy: [https://github.com/meta-llama/llama3](https://github.com/meta-llama/llama3)). This section presents the experimental results of merging the delta parameters obtained through SFT and DPO using the LoRA technique. We also study another preference alignment method ORPO for PAFT, which results in the same observations and conclusions as those from DPO. It shows the generalizability of PAFT to different preference alignment techniques. Due to space limit, we put the experimental results for ORPO in the appendix.\\n\\nFollowing the Zephyr work Tunstall et al. (), we use the UltraChat Ding et al. () dataset for SFT and the UltraFeedback Tunstall et al. () dataset for DPO. UltraChat is a self-refinement dataset consisting of 200K multi-turn dialogues generated by GPT-3.5-Turbo over 30 topics and 20 different types of text material. UltraFeedback consists of 64k prompts, each of which have four LLM\\nresponses that are rated by GPT-4 according to criteria like instruction-following, honesty, and helpfulness.\\n\\nWe meticulously explore a spectrum of merging methods, including SLERP, Task Arithmetic, TIES, DARE-enhanced TIES, and Linear combination. Each of these merging strategies is scrutinized to determine its efficacy in integrating the sparsity-induced parameters from LoRA with the original pre-trained models. The goal is to ascertain which method most effectively preserves the performance enhancements attributed to SFT and DPO, thereby contributing to the advancement of model merging methods in LLM research. For training individual adapters, we have used the same settings as in the *zephyr-7b-beta* development(^3^33[https://github.com/huggingface/alignment-handbook/tree/main/recipes/zephyr-7b-beta](https://github.com/huggingface/alignment-handbook/tree/main/recipes/zephyr-7b-beta)). Our evaluation is conducted using the EleutherAI’s LM Evaluation Harness framework Gao et al. (). We adhere to the same branch (b281b09) used by the HuggingFace Open LLM Leaderboard Beeching et al. (), and evals are run with batch size 1 on an A100 GPU.\\n\\nThe hyper parameter $\\\\lambda$ in Equation 1 controls the sparsity of $\\\\delta_{\\\\text{sft}}$. Empirical values $0.0001$ and $0.001$ are validated in our experiments to achieve reasonable sparsity.\\n\\n**Table 1: Results of compared methods on the six benchmark tasks**\\n| Base Model: Mistral-7B-v0.1 | | | | | | | |\\n| --- | --- | --- | --- | --- | --- | --- | --- |\\n| Method | ARC | HellaSwag | MMLU | TruthfulQA | Winograde | GSM8K | *AVERAGE* |\\n| PAFT ($\\\\text{SFT}_{\\\\text{sparse}}$+DPO) | | | | | | | |\\n| SLERP | 0.6391 | 0.8464 | 0.63961 | 0.5123 | 0.794 | 0.4223 | 0.64228 |\\n| Task Arithmetic | 0.6519 | 0.8477 | 0.63325 | 0.563 | 0.794 | 0.4071 | 0.64949 |\\n| TIES | 0.6519 | 0.8551 | 0.63927 | 0.5453 | 0.7946 | 0.4284 | 0.65243 |\\n| DARE TIES | 0.6493 | 0.8526 | 0.63444 | 0.5454 | 0.7964 | 0.4094 | 0.64792 |\\n| Linear | 0.6348 | 0.8451 | 0.64275 | 0.505 | 0.7932 | 0.4246 | 0.64091 |\\n| Parallel SFT+DPO | | | | | | | |\\n| SLERP | 0.6391 | 0.8479 | 0.63937 | 0.5031 | 0.7924 | 0.4124 | 0.63904 |\\n| Task Arithmetic | 0.651 | 0.851 | 0.62998 | 0.5397 | 0.8011 | 0.4117 | 0.64741 |\\n| TIES | 0.5956 | 0.8319 | 0.61651 | 0.3993 | 0.7853 | 0.3071 | 0.58928 |\\n| DARE TIES | 0.5922 | 0.8244 | 0.60471 | 0.3801 | 0.7577 | 0.2767 | 0.57263 |\\n| Linear | 0.6391 | 0.846 | 0.63935 | 0.4946 | 0.7995 | 0.4314 | 0.64166 |\\n| Sequential | | | | | | | |\\n| $\\\\text{SFT}_{\\\\text{sparse}}$+DPO | 0.6391 | 0.8464 | 0.63461 | 0.5103 | 0.7894 | 0.4123 | 0.63868 |\\n| SFT+DPO | 0.656 | 0.8459 | 0.62634 | 0.5079 | 0.7884 | 0.3836 | 0.63469 |\\n| Individual | | | | | | | |\\n| $\\\\text{SFT}_{\\\\text{sparse}}$-alone | 0.6126 | 0.8233 | 0.6421 | 0.4124 | 0.7711 | 0.3715 | 0.6055 |\\n| SFT-alone | 0.6101 | 0.8216 | 0.6263 | 0.4486 | 0.7798 | 0.3525 | 0.6065 |\\n| DPO-alone | 0.6314 | 0.8487 | 0.6423 | 0.4496 | 0.7932 | 0.4344 | 0.6333 |\\n| Mistral-7B-v0.1 | 0.6049 | 0.8320 | 0.6369 | 0.4259 | 0.7814 | 0.37 | 0.6085 |\\n| Base Model: Llama-3-8B | | | | | | | |\\n| Method | ARC | HellaSwag | MMLU | TruthfulQA | Winograde | GSM8K | *AVERAGE* |\\n| PAFT ($\\\\text{SFT}_{\\\\text{sparse}}$+DPO) | | | | | | | |\\n| SLERP | 0.6067 | 0.8367 | 0.66995 | 0.5297 | 0.7837 | 0.5095 | 0.65604 |\\n| Task Arithmetic | 0.6118 | 0.8411 | 0.66858 | 0.5552 | 0.7806 | 0.5208 | 0.66301 |\\n| TIES | 0.6101 | 0.8414 | 0.67098 | 0.5313 | 0.7891 | 0.5185 | 0.66023 |\\n| DARE TIES | 0.6067 | 0.8398 | 0.66945 | 0.5232 | 0.7885 | 0.5163 | 0.65732 |\\n| Linear | 0.6049 | 0.8329 | 0.67059 | 0.5168 | 0.7837 | 0.5011 | 0.65166 |\\n| Parallel SFT+DPO | | | | | | | |\\n| SLERP | 0.6152 | 0.8347 | 0.66248 | 0.5149 | 0.7869 | 0.5171 | 0.65521 |\\n| Task Arithmetic | 0.6254 | 0.837 | 0.66089 | 0.5266 | 0.7869 | 0.5133 | 0.65835 |\\n| TIES | 0.5879 | 0.8092 | 0.65863 | 0.4283 | 0.7545 | 0.4291 | 0.61127 |\\n| DARE TIES | 0.6007 | 0.8061 | 0.65702 | 0.4233 | 0.7609 | 0.4049 | 0.60882 |\\n| Linear | 0.6152 | 0.8331 | 0.66614 | 0.5082 | 0.7845 | 0.5095 | 0.65277 |\\n| Sequential | | | | | | | |\\n| $\\\\text{SFT}_{\\\\text{sparse}}$+DPO | 0.5648 | 0.7984 | 0.62204 | 0.4049 | 0.7766 | 0.3692 | 0.58932 |\\n| SFT+DPO | 0.5623 | 0.7976 | 0.62258 | 0.4057 | 0.7719 | 0.3662 | 0.58771 |\\n| Individual | | | | | | | |\\n| $\\\\text{SFT}_{\\\\text{sparse}}$-alone | 0.5862 | 0.8177 | 0.66328 | 0.4834 | 0.7719 | 0.4473 | 0.6283 |\\n| SFT-alone | 0.6084 | 0.8135 | 0.65325 | 0.4469 | 0.7648 | 0.4637 | 0.62509 |\\n| DPO-alone | 0.6152 | 0.8412 | 0.6682 | 0.5273 | 0.7845 | 0.4849 | 0.65355 |\\n| Llama-3-8B | 0.5947 | 0.8209 | 0.66603 | 0.4391 | 0.7719 | 0.4587 | 0.62522 |\\n\\n### 3.2 Parallel Training vs. Sequential Training\\n\\nTo demonstrate the advantages of parallel training PAFT, we conducted empirical comparison of parallel, sequential and standalone training approaches on the six benchmark tasks using the two pre-trained models: Mistral-7B and Llama-3-8B. The results are given in Table 1. In the Mistral-7B model section, training with DPO alone improves the average score over the base model, while training with SFT alone doesn’t show an improvement. This result reveals that SFT, while focusing on downstream tasks, inadvertently undermines performance due to a lack of alignment with human preferences. Conversely, DPO aims to harmonize the outputs of LLMs with human preferences, resulting in a noticeable improvement in the average score.\\n\\nFurthermore, we evaluated the sequential training of SFT with L1 regularization followed by DPO, which gave an average score of 0.6387. This score marginally surpasses that of standalone DPO, setting the stage for a comparison with parallel training outcomes. This outcome aligns with our initial hypothesis that during the DPO phase the model appears to discard much of the knowledge acquired in the SFT stage, i.e., alignment tax. Consequently, its performance exhibits only a marginal improvement over the training with DPO-alone.\\n\\nAdditionally, we performed side-by-side evaluations of $\\\\text{SFT}_{\\\\text{sparse}}$+DPO training in both parallel and sequential manners. The findings indicate that training SFT with L1 regularization alongside DPO in parallel leads to a performance metric of 0.6524 when merging with the TIES method, over 2% higher than the score achieved by either DPO alone or by training $\\\\text{SFT}_{\\\\text{sparse}}$ and DPO in sequence. This outcome can be explained by a notable drawback of sequential training which is its tendency to overlook much of the knowledge gained during the SFT stage, suggesting a suboptimal use of SFT data. In contrast, parallel training effectively combines the benefits from SFT and DPO by processing them concurrently. The benefits are mostly preserved during model merging, ensuring efficient utilization of both SFT and DPO data. Our work underscores the enhanced efficacy of the parallel training approach PAFT, which not only maintains the distinct advantages of SFT and DPO, but also outperforms these techniques when they are used separately or sequentially.\\n\\n### 3.3 Sparse Merging vs. Dense Merging\\n\\nOur study has demonstrated the advantages of incorporating sparsity into fine-tuned models. In the context of sequential training, the inclusion of L1 regularization has yielded a modest yet notable improvement. Specifically, in Table 1, the average score for the sequential $\\\\text{SFT}_{\\\\text{sparse}}$+DPO stands at 0.6387, surpassing the sequential SFT+DPO without L1 regularization, with a score of 0.6347. Although the improvement is marginal, it underscores the value of integrating the L1-norm to induce sparsity.\\n\\n**Table 2: Comparison with state-of-the-art LLMs on Open LLM Leaderboard (All the scores are obtained from the Leaderboard)**\\n| LLM | ARC | HellaSwag | MMLU | TruthfulQA | Winograde | GSM8K | *AVERAGE* |\\n| --- | --- | --- | --- | --- | --- | --- | --- |\\n| PAFT (Ein-70B) | 0.7986 | 0.9149 | 0.7805 | 0.7514 | 0.8777 | 0.7544 | 0.8129 |\\n| Mixtral-8x22B-Instruct | 0.727 | 0.8908 | 0.7777 | 0.6814 | 0.8516 | 0.8203 | 0.7915 |\\n| Llama-3-70B-Instruct | 0.7142 | 0.8569 | 0.8006 | 0.6181 | 0.8287 | 0.8544 | 0.7788 |\\n| PAFT (TextBase-7B) | 0.7389 | 0.9027 | 0.6478 | 0.7813 | 0.8603 | 0.6793 | 0.7684 |\\n| Cohere-Command-R+ | 0.7099 | 0.8856 | 0.7573 | 0.563 | 0.854 | 0.7074 | 0.7462 |\\n| DBRX-132B-Instruct | 0.6783 | 0.8885 | 0.7372 | 0.6702 | 0.8208 | 0.6732 | 0.7447 |\\n| OpenChat-3.5 | 0.6604 | 0.8293 | 0.6504 | 0.519 | 0.8177 | 0.6816 | 0.693 |\\n| Llama-3-8B-Instruct | 0.6075 | 0.7855 | 0.6707 | 0.5165 | 0.7451 | 0.6869 | 0.6687 |\\n| Mistral-7B-Instruct-v0.2 | 0.6314 | 0.8488 | 0.6078 | 0.6826 | 0.7719 | 0.4003 | 0.6571 |\\n| Gemma-7B | 0.6109 | 0.8247 | 0.6603 | 0.4491 | 0.7845 | 0.5277 | 0.6429 |\\n\\n**Table 3: Comparison with state-of-the-art LLMs on the AlpacaEval benchmark using GPT-4 as a judge**\\n| LLM | LC WinRate | WinRate |\\n| --- | --- | --- |\\n| GPT-4 Preview | 50.0% | 50.0% |\\n| Claude 3 Opus | 40.5% | 29.1% |\\n| PAFT 70B | 38.6% | 26.5% |\\n| GPT-4 (03/14) | 35.3% | 22.1% |\\n| Claude 3 Sonnet | 34.9% | 25.6% |\\n| Llama 3 70B Instruct | 34.4% | 33.2% |\\n| Mixtral 8x22B v0.1 | 30.9% | 22.2% |\\n| PAFT 7B | 30.6% | 22.8% |\\n| DBRX Instruct | 25.4% | 18.4% |\\n| Mixtral 8x7B v0.1 | 23.7% | 18.3% |\\n| Llama 3 8B Instruct | 22.9% | 22.6% |\\n| GPT 3.5 Turbo | 22.7% | 14.1% |\\n| Mistral 7B v0.2 | 17.1% | 14.7% |\\n\\nThe impact of sparsity becomes more pronounced when examining parallel training scenarios. Across all considered model merging techniques, Parallel $\\\\text{SFT}_{\\\\text{sparse}}$+DPO, i.e., PAFT, consistently outperforms its counterpart without L1 regularization, Parallel SFT+DPO, thereby highlighting the efficacy of the sparsity induced by L1-norm. Notably, in the case of the TIES and DARE TIES merging methods, the average score disparity is significant. With TIES, PAFT ($\\\\text{SFT}_{\\\\text{sparse}}$+DPO) achieves a score of 0.6524, while Parallel SFT+DPO without sparsification lags behind at 0.5893. Similarly, for DARE TIES, PAFT ($\\\\text{SFT}_{\\\\text{sparse}}$+DPO) scores 0.6479, outstripping Parallel SFT+DPO’s 0.5726. This substantial margin illustrates the robustness of L1-norm sparsity for various merging methods.\\n\\nThe same insights as given in the Mistral-7B section can be gained from the Llama-3-8B section in Table 1. PAFT on Llama-3-8B significantly outperforms Parallel SFT+DPO, sequential training and standalone training. The experimental results confirm the generalizability of PAFT to various pre-trained models.\\n\\nWhen comparing different model merging strategies, TIES generally performs better than other methods on both Mistral-7B and Llama-3-8B, exhibiting superior performance over DARE TIES. DARE, which stands for \\\"Drop And REscale\\\", is a method that explicitly increases sparsity by eliminating elements below a certain threshold and rescaling the remaining parameters. In contrast, the L1-norm introduces sparsity implicitly by integrating it into the objective function. Consequently, the impact of the eliminated terms is less pronounced in the final results compared to DARE. This comparison reveals the advantages of the L1-norm’s explicit sparsity induction over the implicit approach employed by DARE.\\n\\n### 3.4 Comparison with State-of-the-art LLMs\\n\\nOn the online Open LLM Leaderboard, we performed PAFT on the Neurotic-7B(^4^44[https://huggingface.co/liminerity/Neurotic-Jomainotrik-7b-slerp](https://huggingface.co/liminerity/Neurotic-Jomainotrik-7b-slerp)) and MoMo-70B(^5^55[https://huggingface.co/leejunhyeok/MoMo-70B-LoRA-V1.2_1](https://huggingface.co/leejunhyeok/MoMo-70B-LoRA-V1.2_1)) base models. The two PAFT-ed models significantly improved over the respective base models, and achieved Rank #1 in the 7B/8B model category and globally on the online Open LLM Leaderboard, respectively, showing the effectiveness of PAFT on various base models. Table 2 gives the results of our PAFT-ed models and the existing state-of-the-art models on the Leaderboard.\\n\\nAdditionally, we compared the two PAFT-ed models with existing state-of-the-art LLMs on the AlpacaEval benchmark Li et al. (), where every model generates responses to 805 questions on different topics, mostly focused on helpfulness. The models are judged by GPT-4, and the final metric is the pairwise win-rate against GPT-4. As shown in Table 3, the PAFT-ed 70B model outperforms existing state-of-the-art LLMs, except *GPT-4 Preview* and *Claude 3 Opus* in LC (Length-controlled) Win-Rate. While the GPT-4 judge favors its own GPT model family, the PAFT-ed 70B model performs better than *GPT-4 (03/14)* and *GPT 3.5 Turbo* do. On the other hand, the PAFT-ed 7B model outperforms all the 7B/8B and smaller models on AlpacaEval. It even beats some larger models, such as *DBRX Instruct* and *Mixtral 8x7B*.\\n\\n## 4 Related Work\\n\\n### 4.1 SFT and Human Preference Alignment\\n\\nThe groundbreaking achievements of BERT Devlin et al. () and GPT OpenAI () have underscored the significance of pretraining and supervised fine-tuning (SFT) techniques. To mitigate ethical concerns and ensure such language model outputs are aligned with human values, a subsequent alignment step employs human feedback to enhance the efficacy of pretraining Christiano et al. (), fine-tuning Ziegler et al. (), and adaptability for scaling purposes Leike et al. (). Kreutzer et al. () found that implicit task feedback often outperforms explicit user feedback, leading to other high-quality datasets of human-generated summaries to compare with those produced by LLMs, resulting in superior quality outputs compared to SFT and human benchmarks Stiennon et al. (). Recent advancements by models such as GPT OpenAI (), Claude Bai et al. (), Llama Touvron et al. (), and Gemini Team () have all leveraged human comparison feedback to refine output quality through alignment, a method also known as reinforcement learning from human feedback (RLHF).\\n\\nRLHF models employ the Bradley-Terry model to develop a reward function that emulates human preferences between two candidate responses Bradley and Terry (). This reward model lays the groundwork for applying reinforcement learning to LLMs, drawing inspiration from Proximal Policy Optimization (PPO) techniques Schulman et al. (). Direct Preference Optimization (DPO) streamlines the alignment process by integrating reward training with LLM alignment, thereby simplifying the training regimen through a direct relationship between the reward function and policy in reinforcement learning Rafailov et al. (). However, the efficacy of DPO in practice remains an area for further exploration Xu et al. (). Odds-ratio Preference Optimization (ORPO) Hong et al. () is an alternative alignment paradigm that aims to replace sequential SFT + DPO with a single monolithic optimization algorithm. It directly optimizes for preferences between two candidate generations by maximizing the ratio of odds of the winning generation w.r.t. losing generation to simultaneously reward logits of desired tokens and penalize logits of undesired tokens.\\n\\nSFT and Human Preference Alignment serve distinct objectives and should be approached as components of a multi-objective optimization problem. SFT focused on enhancing the performance of LLMs in downstream tasks, whereas alignment seeks to address ethical concerns. Prior research on RLHF often treats alignment as a compromise that could potentially degrade the model’s output quality while address ethical problems Ouyang et al. (). Consequently, SFT and alignment are typically implemented in a sequential manner to ensure the safety of LLMs while accepting some degree of capability loss Hou et al. (). In contrast, Bai et al. have claimed that ’Smaller models experience severe ‘alignment taxes’ – their performance on a wide variety of evaluations declines after RLHF training. However, we find a variety of alignment bonuses, with our 13B and 52B RLHF-trained models performing better at zero-shot NLP evaluations, and the same at few-shot evaluations’ Bai et al. (). This divergence in findings motivates further exploration into the interplay between SFT and alignment. Specifically, there is a strong interest in devising a method to integrate SFT and alignment in such a manner that yields an ’alignment bonus.’\\n\\n### 4.2 Sparsity for LLMs\\n\\nAs the size of LLMs continues to increase, the importance of compression becomes crucial for deploying them on edge devices. This is done to reduce costs and improve inference speed Zhu et al. (). Various compression strategies for LLMs exist, with a focus on pruning Han et al. () and Low Rank Adapters (LoRA) Hu et al. (). Pruning involves creating sparsity through pretraining, magnitude-based pruning, and fine-tuning the remaining weights Han et al. (). LoRA suggests representing a matrix as the product of two low-rank matrices to reduce memory storage requirements Hu et al. (). Recent research has shown that the magnitudes of parameters trained by LoRA in SFT process are relatively small. A strategy has been developed where random pruning is applied to these small SFT parameters with a ratio $p$, followed by multiplying the remaining parameters by $\\\\frac{1}{1-p}$ to enhance model performance Yu et al. (). Merging sparsity models trained on different tasks has led to significant improvements in downstream tasks like AlpacaEval and GSM8K. This method involves applying pruning to introduce more sparsity in SFT using LoRA. Other methods for inducing sparsity in SFT parameters exist like incorporating the L1 norm in the loss function, similar to techniques used in Lasso regression Santosa and Symes () and compressed sensing Candes et al. (). A Bayesian interpretation of the L1-norm on the weights amounts to assuming a standard Laplacian prior on the parameters which is centered more closely around mean of zero. This concept will guide the research in this paper.\\n\\n### 4.3 Model Merging\\n\\nCombining skills learnt from different types of datasets in a single model provides multiple benefits like better in-domain performance Poth et al. (), out-of-domain generalization Wang et al. (), and a more parameter efficient model w.r.t. specialized models. Joint multi-task learning is one way to achieve this, but it has several difficulties: it is costly to train a single model across all tasks and it is non-trivial to find the correct task-mix to ensure a jointly optimal performance across all tasks Fifty et al. (). A wide variety of model merging methods to combine specialized models into a stronger merged model have emerged as an alternative to multi-task training. Wortsman et al. () introduced the paradigm of averaging model weights from separate fine-tuned models to create a stronger merged model in ModelSoup, achieving SOTA in several different benchmarks. Fisher merging from Matena and Raffel () proposed to improve upon naively averaging all model weights by instead using a weighted average of the parameters. They identified the importance of each individual parameter based on its Fisher Information to use as the coefficient in the weighted average. Ilharco et al. () further showed that one could influence the merged model’s performance in several ways via task-arithmetic on task-vectors (additive weight adaptors): forgetting undesired traits via negation, learning tasks by addition, or learning entirely new tasks by analogies. Jin et al. () proposed RegMean where they solve a local closed-form linear-regression problem to estimate the merged model parameters for each individual linear layer. Yadav et al. () demonstrated that the phenomenon of parameter interference during model-merging leads to performance degradation in merged models. They cited this interference to two main sources - redundant parameter-updates, i.e. updates not crucial to a model’s prediction, and sign disagreement between different parameter-updates. To overcome such destructive interference, they proposed TIES-Merging which has two filtering steps before model-merging. First, only the top-k% updates by magnitude are retained in each task-vector. Next, the dominant sign is chosen as $\\\\text{sgn}(\\\\Sigma_{i}(\\\\text{sgn}(\\\\theta_{i})))$ and only those updates whose sign agrees with the dominant sign are finally averaged and merged.\\n\\n## 5 Conclusions\\n\\nLLM fine-tuning generally undergoes a two-stage training process, with SFT applied initially, followed by preference alignment. Yet, research indicates that this sequential approach incurs an \\\"alignment tax\\\", compromising the LLM’s overall performance. To counteract this, we advocate for a parallel training strategy PAFT which preserves the advantages of both SFT and preference alignment without incurring the alignment tax associated with sequential training. A significant hurdle in parallel training is the potential for conflict during the model merging phase, where the merging of different adapters can lead to diminished performance. In this paper, we propose the integration of an L1 regularization to the training loss during the SFT phase to induce sparsity, thereby reducing interference between models.\\n\\nOur experimental results demonstrate the efficacy of incorporating an L1-norm into the SFT process for sparsification and utilizing a parallel training framework over the typical sequential approach. When combining all of them together, i.e. Parallel $\\\\text{SFT}_{\\\\text{sparse}}$+DPO achieves the state-of-art results on both the LLM leaderboard by HuggingFace and the AlpacaEval benchmark. The ORPO experimental results given in the appendix show the same patterns, demonstrating the generalizability of our PAFT to various preference alignment methods. This comprehensive strategy highlights how the methods of integrating SFT with preference alignment can greatly enhance LLM fine-tuning.\\n\\n## 6 Limitations\\n\\nThere are a couple of limitations of the parallel training of SFT and preference alignment.\\nFirstly, we have found that sparsity aids in model merging, though the reasons behind this benefit and why DPO initially induces sparsity in the adapter remain unanswered.\\n\\nMoreover, sparsity can reduce model interference during merging, but the scalability of this approach is still in question. If a merged model deployed in production fails in some cases, it is underexplored how to improve the model responses in these cases. Directly performing SFT on the merged model may lead to catastrophic forgetting of what it learned earlier. On the other hand, parallel training necessitates merging a new SFT-ed model with the existing merged model, adding complexity to the process.\\n\\nThe primary risk associated with this paper pertains to its data usage. Currently, UltraChat data is employed for SFT, while UltraFeedback data is used for preference alignment. UltraChat consists solely of multi-round dialogue data, which inherently limits its format diversity. To enhance the robustness and applicability of the model, it is crucial to incorporate a wider variety of data types beyond dialogue data. Additionally, UltraFeedback relies on annotations generated by GPT-4, which inevitably include errors and in-accurate feedback. To mitigate these risks, higher-quality datasets are needed in the future.\\n\\n## Appendix A PAFT Performance with a Different Preference Optimization Algorithm\\n\\nThe stronger performance of PAFT is also confirmed with a different choice of preference alignment algorithm. Table 4 shows experimental results with ORPO as the preference alignment method alongside SFT with the Llama-3-8B base model. We observe a similar trend where finetuning the LLM sequentially via SFT followed by ORPO underperforms all the parallelly trained variants. Even simple model merging methods such as Task Arithmetic and Linear merging perform strongly, outperforming more complicated methods like DARE TIES in both experiment settings.\\n\\n**Table 4: Results of compared methods with ORPO on the six benchmark tasks**\\n| Base Model: Meta-Llama-3-8B | | | | | | | |\\n| --- | --- | --- | --- | --- | --- | --- | --- |\\n| Method | ARC | HellaSwag | MMLU | TruthfulQA | Winograde | GSM8K | *AVERAGE* |\\n| PAFT ($\\\\text{SFT}_{\\\\text{sparse}}$+ORPO) | | | | | | | |\\n| SLERP | 0.599 | 0.8217 | 0.665 | 0.4926 | 0.7845 | 0.4898 | 0.6421 |\\n| Task Arithmetic | 0.5964 | 0.8214 | 0.6655 | 0.4995 | 0.783 | 0.4814 | 0.6412 |\\n| TIES | 0.5947 | 0.8226 | 0.66358 | 0.4931 | 0.783 | 0.4852 | 0.64036 |\\n| DARE TIES | 0.593 | 0.8224 | 0.6637 | 0.4921 | 0.783 | 0.4738 | 0.638 |\\n| Linear | 0.5964 | 0.8206 | 0.6654 | 0.4923 | 0.7814 | 0.4905 | 0.6411 |\\n| Parallel SFT+ORPO | | | | | | | |\\n| SLERP | 0.6049 | 0.8227 | 0.668 | 0.4905 | 0.783 | 0.4951 | 0.644 |\\n| Task Arithmetic | 0.6152 | 0.8209 | 0.6621 | 0.4908 | 0.7845 | 0.4989 | 0.6454 |\\n| TIES | 0.593 | 0.8139 | 0.6633 | 0.4446 | 0.768 | 0.467 | 0.6250 |\\n| DARE TIES | 0.5981 | 0.8101 | 0.66 | 0.4398 | 0.7632 | 0.4534 | 0.6208 |\\n| Linear | 0.6067 | 0.8222 | 0.6685 | 0.4868 | 0.783 | 0.4989 | 0.6444 |\\n| Sequential | | | | | | | |\\n| $\\\\text{SFT}_{\\\\text{sparse}}$+ORPO | 0.5563 | 0.8018 | 0.62116 | 0.4068 | 0.7719 | 0.3662 | 0.58736 |\\n| SFT+ORPO | 0.5589 | 0.8021 | 0.62142 | 0.4092 | 0.7711 | 0.3677 | 0.5884 |\\n| Llama-3-8B | 0.5947 | 0.8209 | 0.64854 | 0.4391 | 0.7719 | 0.4587 | 0.62231 |\",\n \"metadata\": {\n \"authors\": [\n \"Shiva Kumar Pentyala\",\n \"Zhichao Wang\",\n \"Bin Bi\",\n \"Kiran Ramnath\",\n \"Xiang-Bo Mao\",\n \"Regunathan Radhakrishnan\",\n \"Sitaram Asur\",\n \"Na (Claire) Cheng\"\n ],\n \"title\": \"PAFT: A Parallel Training Paradigm for Effective LLM Fine-Tuning\",\n \"total_chars\": 39326,\n \"total_lines\": 231,\n \"updated_at\": \"2026-01-25T14:38:43.390663\"\n }\n}"},"reviews":{"kind":"list like","value":[{"id":"0jDZzZKhOK","initial_rating":5,"confidence":4,"soundness":2,"contribution":3,"presentation":4,"summary":"This paper proposes a novel training approach to perform parallel training of SFT and preference alignment (like PPO, DPO). Their approach aims at solving the problem (e.g., catastrophic forgetting) induced by the traditional sequential training approach (first SFT and then preference alignment).","strengths":"1. The idea of this paper is interesting and worth being investigated.\n2. The proposed method seems to be effective according to the provided results.","weaknesses":"1. **Unclear motivation towards the inclusion of adapters in the proposed method.**\n- The authors should explain why adapters are needed to address the sequential training problem of SFT and preference alignment.\n2. **Vague properties abut the proposed approach.** \n- Deeper analysis about the following items should have been included in the paper.\n - The method design\n - Why the proposed method works\n - Improvements over catastrophic forgetting\n3. **Limited novelty of the proposed method.**\n- The sparse merging actually rely on existing methods.\n- The contributions will increase if a customized merging approach is proposed for bridging SFT and preference alignment.\n4. Poor formatting of Table 1 and Table 2 (can be fixed).","questions":"1. How was sparsity discoverd to affect merging process?\n2. What do \"Ein-70B\" and \"TextBase-7B\" represent in Table 2?\n3. The authors performed PAFT on the Neurotic-7B 4 and MoMo-70B models. However, Table 3 does not include the original scores of Neurotic-7B 4 and MoMo-70B before PAFT. Is there a reason behind this?\n4. Can the proposed approach work without adapters?"},{"id":"ckdvA8CKOh","initial_rating":3,"confidence":4,"soundness":3,"contribution":2,"presentation":2,"summary":"Large language models (LLMs) often go through a sequential training pipeline before deployment, i.e., supervised finetuning (SFT) followed by preference alignment, such as DPO. However, the alignment could lead to performance reduction on some capabilities acquired during pretraining and SFT, namely the alignment tax. This paper addresses this problem by replacing the sequential pipeline with a parallel training pipeline, namely PAFT: it performs SFT and DPO independently based on the same pretrained base model, and then merges them as the final model. The authors observed that increasing sparsity to SFT leads to improved PAFT performance. Through experiments with Mistral and LLama, the authors show that PAFT outperforms sequential training, and achieves top results in Huggingface Open LLM leaderboard.","strengths":"* PAFT represents a new and interesting post-training paradigm, that performs SFT and DPO independently and gets the final model by merging them;\n* PAFT shows improved performance compared to individual SFT/DPO and sequential training;\n* Experiments are based on two different LLM families;","weaknesses":"* Lack of explanation why pretrained model can be directly used for alignment and leads to excellent performance;\n* Experimental setup doesn’t support alignment tax, but SFT tax;\n* Evaluation on alignment performance is weak but important;\n* Statements are not always supportive and require more experiments","questions":"1. While independent DPO and SFT sounds interesting, it’s unclear why a pretrained model could be directly used for DPO or alignment. In the literature and in practice, SFT is often used before alignment to make sure the model could produce reasonable responses. More explanation should be given.\n2. In Table 1, it’s surprising that DPO alone outperforms SFT and base model by a large margin, which means the alignment doesn’t suffer from tax but improves downstream tasks consistently. This is counter-intuitive and the authors need to provide more convincing analysis.\n3. Besides, Table 1 shows that SFT hurts downstream performance. The authors may need to consider changing their experimental setup. Or give the readers more explanations.\n4. Major experiments are based on performance on six downstream tasks, which however has little to do with human preferences. It’s unclear how the proposed method affects preference alignment compared to the baselines given in Table 1. Table 3 is the right evaluation direction but no comparisons with the baselines.\n5. In the paper, the authors argue that sparsity is an important factor but didn’t show further ablations. What if applying L1 norm to DPO? Would it lead to better performance? What if reducing \\lambda in Equation (1) to produce SFT models at different levels of sparsity? Would higher sparsity really lead to better results?\n6. In Table 1, SFT-alone and DPO-alone achieve an average score of 60.65 and 63.33 respectively, but the merged model achieves a score of 65.24. Could you provide a simple interpretation where the improvements come from?"},{"id":"o7yRJeDVwv","initial_rating":8,"confidence":4,"soundness":3,"contribution":4,"presentation":3,"summary":"The authors propose PAFT (Parallel Fine-Tuning for LLMs), a novel training paradigm that fine-tunes large language models (LLMs) by concurrently applying supervised fine-tuning (SFT) and preference alignment rather than sequentially. This parallel approach minimizes the \"alignment tax\"—the degradation of previously fine-tuned knowledge when preference alignment follows fine-tuning. PAFT further introduces sparsification of the SFT delta parameters using L1 regularization, enhancing the compatibility of both fine-tuning phases during parameter merging. Experimental results show that PAFT outperforms sequential methods, achieving high rankings on benchmarks like the HuggingFace Open LLM Leaderboard and AlpacaEval.","strengths":"Effective Mitigation of Alignment Tax: PAFT introduces a parallel training paradigm that addresses catastrophic forgetting often caused by sequential fine-tuning. The approach preserves fine-tuning and alignment benefits without performance compromise.\nInnovative Sparsification Technique: By applying L1 regularization to the SFT delta parameters, PAFT effectively reduces parameter interference during model merging, resulting in a more efficient and robust model.\nComprehensive Benchmarking: The authors demonstrate the effectiveness of PAFT across multiple datasets and merging methods, establishing its robustness and achieving top performance across multiple tasks.\nState-of-the-Art Results: PAFT achieved Rank #1 for specific model sizes on the HuggingFace Leaderboard, demonstrating practical impact and the potential for real-world deployment.","weaknesses":"Sparse Experimental Analysis on Broader Tasks: While PAFT’s strengths are demonstrated for tasks in NLP benchmarks, additional experiments on diverse domains would enhance the generalizability claims.\nLimited Baseline Comparison: The study primarily compares PAFT to sequential and standalone training, with minimal inclusion of alternative parallel training strategies. Introducing comparisons with other recent sparse training or parallel tuning techniques could contextualize PAFT's advantages.\nPotential Overhead in Parallel Execution: Implementing parallel SFT and preference alignment could introduce additional computational overhead. An analysis of training efficiency, including any potential increased resource requirements, would be beneficial.\nIn-depth Ablation of Sparsity Levels: Additional analysis on different L1 regularization strengths would further clarify the trade-offs between sparsity and performance.","questions":"Does the parallel structure of PAFT introduce any latency or computational overhead, particularly in real-time or low-latency applications?\nWhat are the effects of varying L1 regularization strengths on downstream performance and model sparsity? A more detailed ablation study could provide insights into optimal settings.\nHow does PAFT generalize to tasks beyond NLP, such as vision or multi-modal tasks?\nCan other forms of preference alignment, such as ORPO, provide similar improvements? Additional benchmarks using ORPO alongside DPO would strengthen PAFT’s applicability."},{"id":"BSfsHpGYVt","initial_rating":8,"confidence":3,"soundness":3,"contribution":3,"presentation":3,"summary":"1.It shows that parallel training of SFT (Supervised Fine-Tuning) and preference alignment outperforms sequential training, significantly reducing the alignment tax.\n2.The paper emphasizes the importance of sparse model integration to prevent model conflict while maintaining the full functionality of each model. It demonstrates that the L1-norm is superior to DARE (Deep Attentional Reinforcement Learning for Exploration) as a more effective and higher-quality method for promoting sparsity in model training across various model merging techniques.\n3.The paper provides an extensive evaluation of PAFT (Preference Alignment Fine-Tuning) on renowned public benchmarks, such as the Open LLM Leaderboard and AlpacaEval. The PAFT-optimized 7B model secured the #1 rank in the 7B/8B model category on the Open LLM Leaderboard, and the PAFT-optimized 70B model led the global Leaderboard.","strengths":"The paper is well-written, providing a new parallel training paradigm called PAFT for fine-tuning large language models, which to some extent offsets the alignment tax caused by sequential training. The large language model produced by this new training paradigm performs excellently, ranking first in the 7B/8B model category.","weaknesses":"1. There is a lack of brief introductions to the various model merging methods used in the experimental section.\n2. Pay attention to the layout, some tables exceed the width limit, such as Table 1, Table 2, and Table 4.","questions":"1. In Table 1, regardless of whether it is Mistral-7B or Llama-3-8B, the average score of PAFT using the Linear merging method is slightly lower than that of Parallel SFT+DPO, and a similar phenomenon is observed in Table 4. What could be the possible reasons for this?\n2. Why does PAFT show such a significant improvement over Parallel SFT+DPO in the TIES and DARE TIES merging methods, while the improvement is not as noticeable in other model merging methods?"}],"string":"[\n {\n \"id\": \"0jDZzZKhOK\",\n \"initial_rating\": 5,\n \"confidence\": 4,\n \"soundness\": 2,\n \"contribution\": 3,\n \"presentation\": 4,\n \"summary\": \"This paper proposes a novel training approach to perform parallel training of SFT and preference alignment (like PPO, DPO). Their approach aims at solving the problem (e.g., catastrophic forgetting) induced by the traditional sequential training approach (first SFT and then preference alignment).\",\n \"strengths\": \"1. The idea of this paper is interesting and worth being investigated.\\n2. The proposed method seems to be effective according to the provided results.\",\n \"weaknesses\": \"1. **Unclear motivation towards the inclusion of adapters in the proposed method.**\\n- The authors should explain why adapters are needed to address the sequential training problem of SFT and preference alignment.\\n2. **Vague properties abut the proposed approach.** \\n- Deeper analysis about the following items should have been included in the paper.\\n - The method design\\n - Why the proposed method works\\n - Improvements over catastrophic forgetting\\n3. **Limited novelty of the proposed method.**\\n- The sparse merging actually rely on existing methods.\\n- The contributions will increase if a customized merging approach is proposed for bridging SFT and preference alignment.\\n4. Poor formatting of Table 1 and Table 2 (can be fixed).\",\n \"questions\": \"1. How was sparsity discoverd to affect merging process?\\n2. What do \\\"Ein-70B\\\" and \\\"TextBase-7B\\\" represent in Table 2?\\n3. The authors performed PAFT on the Neurotic-7B 4 and MoMo-70B models. However, Table 3 does not include the original scores of Neurotic-7B 4 and MoMo-70B before PAFT. Is there a reason behind this?\\n4. Can the proposed approach work without adapters?\"\n },\n {\n \"id\": \"ckdvA8CKOh\",\n \"initial_rating\": 3,\n \"confidence\": 4,\n \"soundness\": 3,\n \"contribution\": 2,\n \"presentation\": 2,\n \"summary\": \"Large language models (LLMs) often go through a sequential training pipeline before deployment, i.e., supervised finetuning (SFT) followed by preference alignment, such as DPO. However, the alignment could lead to performance reduction on some capabilities acquired during pretraining and SFT, namely the alignment tax. This paper addresses this problem by replacing the sequential pipeline with a parallel training pipeline, namely PAFT: it performs SFT and DPO independently based on the same pretrained base model, and then merges them as the final model. The authors observed that increasing sparsity to SFT leads to improved PAFT performance. Through experiments with Mistral and LLama, the authors show that PAFT outperforms sequential training, and achieves top results in Huggingface Open LLM leaderboard.\",\n \"strengths\": \"* PAFT represents a new and interesting post-training paradigm, that performs SFT and DPO independently and gets the final model by merging them;\\n* PAFT shows improved performance compared to individual SFT/DPO and sequential training;\\n* Experiments are based on two different LLM families;\",\n \"weaknesses\": \"* Lack of explanation why pretrained model can be directly used for alignment and leads to excellent performance;\\n* Experimental setup doesn’t support alignment tax, but SFT tax;\\n* Evaluation on alignment performance is weak but important;\\n* Statements are not always supportive and require more experiments\",\n \"questions\": \"1. While independent DPO and SFT sounds interesting, it’s unclear why a pretrained model could be directly used for DPO or alignment. In the literature and in practice, SFT is often used before alignment to make sure the model could produce reasonable responses. More explanation should be given.\\n2. In Table 1, it’s surprising that DPO alone outperforms SFT and base model by a large margin, which means the alignment doesn’t suffer from tax but improves downstream tasks consistently. This is counter-intuitive and the authors need to provide more convincing analysis.\\n3. Besides, Table 1 shows that SFT hurts downstream performance. The authors may need to consider changing their experimental setup. Or give the readers more explanations.\\n4. Major experiments are based on performance on six downstream tasks, which however has little to do with human preferences. It’s unclear how the proposed method affects preference alignment compared to the baselines given in Table 1. Table 3 is the right evaluation direction but no comparisons with the baselines.\\n5. In the paper, the authors argue that sparsity is an important factor but didn’t show further ablations. What if applying L1 norm to DPO? Would it lead to better performance? What if reducing \\\\lambda in Equation (1) to produce SFT models at different levels of sparsity? Would higher sparsity really lead to better results?\\n6. In Table 1, SFT-alone and DPO-alone achieve an average score of 60.65 and 63.33 respectively, but the merged model achieves a score of 65.24. Could you provide a simple interpretation where the improvements come from?\"\n },\n {\n \"id\": \"o7yRJeDVwv\",\n \"initial_rating\": 8,\n \"confidence\": 4,\n \"soundness\": 3,\n \"contribution\": 4,\n \"presentation\": 3,\n \"summary\": \"The authors propose PAFT (Parallel Fine-Tuning for LLMs), a novel training paradigm that fine-tunes large language models (LLMs) by concurrently applying supervised fine-tuning (SFT) and preference alignment rather than sequentially. This parallel approach minimizes the \\\"alignment tax\\\"—the degradation of previously fine-tuned knowledge when preference alignment follows fine-tuning. PAFT further introduces sparsification of the SFT delta parameters using L1 regularization, enhancing the compatibility of both fine-tuning phases during parameter merging. Experimental results show that PAFT outperforms sequential methods, achieving high rankings on benchmarks like the HuggingFace Open LLM Leaderboard and AlpacaEval.\",\n \"strengths\": \"Effective Mitigation of Alignment Tax: PAFT introduces a parallel training paradigm that addresses catastrophic forgetting often caused by sequential fine-tuning. The approach preserves fine-tuning and alignment benefits without performance compromise.\\nInnovative Sparsification Technique: By applying L1 regularization to the SFT delta parameters, PAFT effectively reduces parameter interference during model merging, resulting in a more efficient and robust model.\\nComprehensive Benchmarking: The authors demonstrate the effectiveness of PAFT across multiple datasets and merging methods, establishing its robustness and achieving top performance across multiple tasks.\\nState-of-the-Art Results: PAFT achieved Rank #1 for specific model sizes on the HuggingFace Leaderboard, demonstrating practical impact and the potential for real-world deployment.\",\n \"weaknesses\": \"Sparse Experimental Analysis on Broader Tasks: While PAFT’s strengths are demonstrated for tasks in NLP benchmarks, additional experiments on diverse domains would enhance the generalizability claims.\\nLimited Baseline Comparison: The study primarily compares PAFT to sequential and standalone training, with minimal inclusion of alternative parallel training strategies. Introducing comparisons with other recent sparse training or parallel tuning techniques could contextualize PAFT's advantages.\\nPotential Overhead in Parallel Execution: Implementing parallel SFT and preference alignment could introduce additional computational overhead. An analysis of training efficiency, including any potential increased resource requirements, would be beneficial.\\nIn-depth Ablation of Sparsity Levels: Additional analysis on different L1 regularization strengths would further clarify the trade-offs between sparsity and performance.\",\n \"questions\": \"Does the parallel structure of PAFT introduce any latency or computational overhead, particularly in real-time or low-latency applications?\\nWhat are the effects of varying L1 regularization strengths on downstream performance and model sparsity? A more detailed ablation study could provide insights into optimal settings.\\nHow does PAFT generalize to tasks beyond NLP, such as vision or multi-modal tasks?\\nCan other forms of preference alignment, such as ORPO, provide similar improvements? Additional benchmarks using ORPO alongside DPO would strengthen PAFT’s applicability.\"\n },\n {\n \"id\": \"BSfsHpGYVt\",\n \"initial_rating\": 8,\n \"confidence\": 3,\n \"soundness\": 3,\n \"contribution\": 3,\n \"presentation\": 3,\n \"summary\": \"1.It shows that parallel training of SFT (Supervised Fine-Tuning) and preference alignment outperforms sequential training, significantly reducing the alignment tax.\\n2.The paper emphasizes the importance of sparse model integration to prevent model conflict while maintaining the full functionality of each model. It demonstrates that the L1-norm is superior to DARE (Deep Attentional Reinforcement Learning for Exploration) as a more effective and higher-quality method for promoting sparsity in model training across various model merging techniques.\\n3.The paper provides an extensive evaluation of PAFT (Preference Alignment Fine-Tuning) on renowned public benchmarks, such as the Open LLM Leaderboard and AlpacaEval. The PAFT-optimized 7B model secured the #1 rank in the 7B/8B model category on the Open LLM Leaderboard, and the PAFT-optimized 70B model led the global Leaderboard.\",\n \"strengths\": \"The paper is well-written, providing a new parallel training paradigm called PAFT for fine-tuning large language models, which to some extent offsets the alignment tax caused by sequential training. The large language model produced by this new training paradigm performs excellently, ranking first in the 7B/8B model category.\",\n \"weaknesses\": \"1. There is a lack of brief introductions to the various model merging methods used in the experimental section.\\n2. Pay attention to the layout, some tables exceed the width limit, such as Table 1, Table 2, and Table 4.\",\n \"questions\": \"1. In Table 1, regardless of whether it is Mistral-7B or Llama-3-8B, the average score of PAFT using the Linear merging method is slightly lower than that of Parallel SFT+DPO, and a similar phenomenon is observed in Table 4. What could be the possible reasons for this?\\n2. Why does PAFT show such a significant improvement over Parallel SFT+DPO in the TIES and DARE TIES merging methods, while the improvement is not as noticeable in other model merging methods?\"\n }\n]","isConversational":false},"scores":{"kind":"string","value":"{\n \"rating\": \"3;5;6;8\",\n \"rating_avg\": 5.5,\n \"confidence\": \"4;4;4;3\",\n \"confidence_avg\": 3.75,\n \"soundness\": \"3;2;3;3\",\n \"soundness_avg\": 2.75,\n \"contribution\": \"2;3;3;3\",\n \"contribution_avg\": 2.75,\n \"presentation\": \"2;4;3;3\",\n \"presentation_avg\": 3\n}"},"metadata":{"kind":"string","value":"{\n \"primary_area\": \"\",\n \"track\": \"main\",\n \"venue\": \"Submitted to ICLR 2025\",\n \"venueid\": \"ICLR.cc/2025/Conference/Rejected_Submission\",\n \"processed_at\": \"2026-01-14T22:16:01.734135\"\n}"},"meta_review":{"kind":"string","value":"{\n \"id\": \"MvyhMh0Z4W\",\n \"metareview\": \"The paper proposes a novel training method for fine-tuning large language models (LLMs). The approach independently trains a model using supervised fine-tuning (SFT) and preference alignment methods like DPO and then merges them using sparsity-aware techniques to address redundancy. The paper claims that the proposed PAFT method mitigates the “alignment tax,” improves downstream task performance, and achieves high rankings on benchmarks like HuggingFace Open LLM Leaderboard. It further highlights that preference alignment like DPO produces inherently sparse models, whereas SFT produces dense models, motivating adding sparsity during SFT to support effective merging. While the paper introduces an interesting hypothesis, it is not well-supported by the experimental evidence and the evaluation methodology, which raises several critical concerns.\\n\\nThe strengths of the paper include an innovative attempt at performing SFT and preference alignment in parallel rather than sequentially, addressing an important problem in LLM alignment. The idea of using sparsity to resolve parameter conflicts in merging is interesting and has potential implications. Additionally, the authors present results across multiple benchmarks, including TruthfulQA, AlpacaEval, and the Open LLM Leaderboard, with the PAFT-optimized models achieving competitive scores. \\n\\nHowever, the paper has several significant weaknesses that ultimately outweigh its strengths. First, there is a fundamental mismatch between the paper's motivation and the results presented. The paper claims to reduce an “alignment tax” but shows that standalone DPO often outperforms SFT and even the sequential combination of SFT and DPO, indicating that the so-called tax stems more from SFT degradation than alignment. This discrepancy undermines the central claim and raises doubts about the problem being addressed. Second, the experimental results lack consistency across merging strategies, and the improvements are minimal or absent in some cases (e.g., Linear merging). Some baselines are missing, including alternative parallel or sparsity-based fine-tuning methods. Furthermore, there is a lack of controlled experiments and comprehensive ablations to convincingly support the role of sparsity in improving performance. DPO’s superior performance when applied directly to pre-trained LLMs contradicts established norms, which typically requires SFT as a prerequisite for meaningful improvements.\\n\\nThe most critical reasons for recommending rejection are the inconsistency between the hypothesis (mitigating alignment tax) and the observed results, insufficient experimental rigor, and a lack of clarity regarding the contributions. While the paper presents a potentially interesting idea, it falls short of providing a solid scientific foundation and controlled evidence to substantiate its claims, limiting its impact and significance. As such, I recommend rejecting the paper.\",\n \"additional_comments\": \"During the rebuttal period, several key issues were raised by reviewers, focusing primarily on the paper’s claims, experimental results, and methodological choices. The main points of contention included: (1) the paper's central claim of mitigating the “alignment tax,” (2) the inconsistent results across baselines and merging methods, (3) insufficient ablations and analysis of sparsity, and (4) the surprising performance of DPO without prior SFT.\\n\\nReviewers MfEm and 2DRE questioned the “alignment tax” claim, pointing out that DPO alone outperforms both SFT and sequential SFT+DPO, which suggests SFT contributes to performance degradation rather than alignment. The authors argued that the datasets and optimization objectives of SFT and DPO differ, making direct comparisons inappropriate. However, this explanation failed to fully convince the reviewers, as it did not reconcile the core inconsistency or provide further analysis to justify the claim.\\n\\nConcerns were raised regarding the experimental results and baselines. Reviewer ShWV noted that improvements using PAFT were not consistent across merging methods, with some approaches (e.g., Linear merging) showing little to no gain over baselines. Additionally, reviewers noted a lack of comparisons with alternative parallel or sparsity-based fine-tuning methods. The authors responded by clarifying their focus on SFT+DPO as the main baseline and by presenting additional results exploring sparsity levels. While these clarifications were helpful, they did not fully resolve the concerns about experimental rigor and completeness.\\n\\nReviewers, including Cf8k, called for deeper ablations on sparsity and its role in PAFT performance. The authors provided further experiments to test different L1 regularization strengths and sparsity levels, showing that λ = 0.001 offered the best performance. While these results added clarity, reviewers remained unconvinced that the sparsity contribution was sufficiently novel or explored across a broad range of settings.\\n\\nLastly, the unexpected performance of DPO alone was flagged by multiple reviewers, especially Reviewer MfEm, who found it counterintuitive that DPO directly applied to a pre-trained LLM outperformed SFT. The authors attributed this to DPO’s preference-based optimization and better generalization. However, this explanation was not fully satisfying, as reviewers suggested further controlled experiments (e.g., fine-tuning SFT on preference datasets) to isolate the cause.\\n\\nIn my decision, I weighed these points as follows: The core claim about mitigating alignment tax remains unsubstantiated due to inconsistent evidence and lack of controlled analysis. While the additional ablations improved the paper, they were limited in scope, and concerns about experimental rigor, baselines, and novelty persisted. The authors' rebuttal clarified several points but did not sufficiently address the central criticisms. Given these unresolved issues, particularly the mismatch between the paper’s motivation and results, I align with the reviewers’ skepticism and recommend rejection.\"\n}"},"decision":{"kind":"string","value":"{\n \"decision\": \"Reject\"\n}"}}}],"truncated":false,"partial":false},"paginationData":{"pageIndex":12,"numItemsPerPage":100,"numTotalItems":4011,"offset":1200,"length":100}},"jwt":"eyJhbGciOiJFZERTQSIsImtpZCI6IjVHZDBvd0g5MTM2eDZjc1FvbE1zcktNWUZoRVFoUm5PVVVybEpjOUhaUEEifQ.eyJyZWFkIjp0cnVlLCJwZXJtaXNzaW9ucyI6eyJyZXBvLmNvbnRlbnQucmVhZCI6dHJ1ZX0sImlhdCI6MTc4NDkwODQwOSwianRpIjoiMDNhMTM2MDEtMTA2YS00M2Q1LTgwNjQtYTEzNmU2ZjdkNWRjIiwic3ViIjoiL2RhdGFzZXRzL1VLUExhYi9Qcm9SZXZpZXdlci1EYXRhc2V0IiwiZXhwIjoxNzg0OTEyMDA5LCJpc3MiOiJodHRwczovL2h1Z2dpbmdmYWNlLmNvIn0.CnH8LQRl8mXq_6xjmIhxL24ptKhPl7efkNv8BSmiT-S6EHMDpqe-eR6iHC0yEI1ldXLtriBT4qD7e3nOnVR4Cg","displayUrls":true,"splitSizeSummaries":[{"config":"default","split":"train","numRows":4011,"numBytesParquet":144344024},{"config":"default","split":"test","numRows":1000,"numBytesParquet":45705060}]},"discussionsStats":{"closed":0,"open":1,"total":1},"fullWidth":true,"hasGatedAccess":true,"hasFullAccess":true,"isEmbedded":false,"savedQueries":{"community":[],"user":[]}}">
paper_id string | arxiv_id string | title string | markdown dict | reviews list | scores dict | metadata dict | meta_review dict | decision dict |
|---|---|---|---|---|---|---|---|---|
IaFAvyaBOc | 2410.03497v1 | Collaborative and Efficient Personalization with Mixtures of Adaptors | {
"content": "## Abstract\n\nAbstract Non-iid data is prevalent in real-world federated learning problems. Data heterogeneity can come in different types in terms of distribution shifts. In this work, we are interested in the heterogeneity that comes from concept shifts, i.e., shifts in the prediction across clients.... | [
{
"id": "0LDy0xVBMN",
"initial_rating": 1,
"confidence": 5,
"soundness": 1,
"contribution": 1,
"presentation": 1,
"summary": "The paper proposes \"FLoRAL,\" a parameter-efficient federated learning framework that uses mixtures of low-rank adaptors (LoRAs) to enable [perhaps] client perso... | {
"rating": "1;3;3;5",
"rating_avg": 3,
"confidence": "5;3;4;4",
"confidence_avg": 4,
"soundness": "1;2;2;2",
"soundness_avg": 1.75,
"contribution": "1;2;1;2",
"contribution_avg": 1.5,
"presentation": "1;1;2;2",
"presentation_avg": 1.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.585147"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
Ian00SaFHg | 2410.04988v1 | Efficient Model-Based Reinforcement Learning Through Optimistic Thompson Sampling | {
"content": "## Abstract\n\nAbstract Learning complex robot behavior through interactions with the environment necessitates principled exploration. Effective strategies should prioritize exploring regions of the state-action space that maximize rewards, with optimistic exploration emerging as a promising direction a... | [
{
"id": "p1t5aZN0uq",
"initial_rating": 8,
"confidence": 3,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "This paper proposed Hallucination-based Optimistic Thompson Sampling with Gaussian Processes (HOT-GP), a method for principled optimistic exploration. It uses a ... | {
"rating": "3;3;6;8",
"rating_avg": 5,
"confidence": "4;3;3;5",
"confidence_avg": 3.75,
"soundness": "2;2;2;4",
"soundness_avg": 2.5,
"contribution": "2;2;2;4",
"contribution_avg": 2.5,
"presentation": "2;3;3;4",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.586394"
} | {
"id": "VcF0JC0uUZ",
"metareview": "The paper proposes HOT-GP, a model-based Thompson sampling-based RL method that takes into account the *joint* uncertainty in dynamics and reward, as modeled by a joint Gaussian Process, when making exploration decisions. The empirical evaluation shows it to be generally more sa... | {
"decision": "Accept (Poster)"
} |
IcMfCFPdd2 | 2409.09984v1 | Convergence of Sharpness-Aware Minimization Algorithms using Increasing Batch Size and Decaying Learning Rate | {
"content": "## Abstract\n\nAbstract The sharpness-aware minimization (SAM) algorithm and its variants, including gap guided SAM (GSAM), have been successful at improving the generalization capability of deep neural network models by finding flat local minima of the empirical loss in training. Meanwhile, it has been... | [
{
"id": "qBxz4mms1c",
"initial_rating": 3,
"confidence": 4,
"soundness": 2,
"contribution": 1,
"presentation": 2,
"summary": "Paper addressen generalization and convergence properties of Deep learning. In particular, incsiperd by previous works, it takes SAM and applies hypothesis of \"i... | {
"rating": "3;3;3;5",
"rating_avg": 3.5,
"confidence": "4;2;4;3",
"confidence_avg": 3.25,
"soundness": "2;2;2;2",
"soundness_avg": 2,
"contribution": "2;2;1;2",
"contribution_avg": 1.75,
"presentation": "1;2;2;3",
"presentation_avg": 2
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.587294"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
IcPkW3QNW2 | 2410.13862v1 | DepthSplat: Connecting Gaussian Splatting and Depth | {
"content": "## Abstract\n\nAbstract Gaussian splatting and single/multi-view depth estimation are typically studied in isolation. In this paper, we present DepthSplat to connect Gaussian splatting and depth estimation and study their interactions. More specifically, we first contribute a robust multi-view depth mod... | [
{
"id": "sEGTX3721f",
"initial_rating": 5,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper introduces DepthSplat, a framework that bridges Gaussian splatting and depth estimation to improve both tasks. By incorporating pre-trained monocular d... | {
"rating": "5;5;5;5;5",
"rating_avg": 5,
"confidence": "4;4;4;4;4",
"confidence_avg": 4,
"soundness": "3;3;3;4;3",
"soundness_avg": 3.2,
"contribution": "3;3;2;2;3",
"contribution_avg": 2.6,
"presentation": "2;3;3;3;3",
"presentation_avg": 2.8
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.588295"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
IcovaKGyMp | 2410.10724v1 | Large Language Models Are Active Critics in NLG Evaluation | {
"content": "## Abstract\n\nAbstract The conventional paradigm of using large language models (LLMs) for evaluating natural language generation (NLG) systems typically relies on two key inputs: (1) a clear definition of the NLG task to be evaluated and (2) a list of pre-defined evaluation criteria. This process trea... | [
{
"id": "18D6zkjUNw",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "When prompting LLMs for NLG evaluation, prompting can be inconvenient due to the required manual intervention to design the prompts, which may also limit the mode... | {
"rating": "3;5;5;5",
"rating_avg": 4.5,
"confidence": "4;4;4;3",
"confidence_avg": 3.75,
"soundness": "2;3;3;3",
"soundness_avg": 2.75,
"contribution": "2;2;2;2",
"contribution_avg": 2,
"presentation": "3;3;4;3",
"presentation_avg": 3.25
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.588940"
} | {
"id": "qIeH2Ne2CF",
"metareview": "After a thorough examination of the paper and the thoughtful feedback provided in the reviewer comments, while I acknowledge this paper addresses an important direction in developing automated NLG evaluation methods with LLMs, there are several significant concerns that need to ... | {
"decision": "Reject"
} |
Idygh9MX0N | 2407.15073v2 | Multi-Agent Causal Discovery Using Large Language Models | {
"content": "## Abstract\n\nAbstract Large Language Models (LLMs) have demonstrated significant potential in causal discovery tasks by utilizing their vast expert knowledge from extensive text corpora. However, the multi-agent capabilities of LLMs in causal discovery remain underexplored. This paper introduces a gen... | [
{
"id": "SQHCbex9CT",
"initial_rating": 1,
"confidence": 3,
"soundness": 1,
"contribution": 2,
"presentation": 1,
"summary": "This paper proposed a meta-debator model for causal discovery using large language models called MAC and perform simulations on three datasets.",
"strengths":... | {
"rating": "1;3;3;5;5",
"rating_avg": 3.4,
"confidence": "3;4;4;3;3",
"confidence_avg": 3.4,
"soundness": "1;3;2;3;3",
"soundness_avg": 2.4,
"contribution": "2;2;1;2;2",
"contribution_avg": 1.8,
"presentation": "1;2;3;2;3",
"presentation_avg": 2.2
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.589578"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
IeRcpsdY7P | 2410.02536v2 | Intelligence at the Edge of Chaos | {
"content": "## Abstract\n\nAbstract We explore the emergence of intelligent behavior in artificial systems by investigating how the complexity of rule-based systems influences the capabilities of models trained to predict these rules. Our study focuses on elementary cellular automata (ECA), simple yet powerful one-... | [
{
"id": "uyjGD84y4j",
"initial_rating": 6,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 2,
"summary": "The paper explores whether LLMs can develop generalisable reasoning abilities by pretraining on structured but simple datasets, specifically sequences generated b... | {
"rating": "6;6;6;6",
"rating_avg": 6,
"confidence": "4;3;3;4",
"confidence_avg": 3.5,
"soundness": "2;3;2;3",
"soundness_avg": 2.5,
"contribution": "3;3;3;3",
"contribution_avg": 3,
"presentation": "3;3;3;2",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.590224"
} | {
"id": "ZFlSboA3ph",
"metareview": "This paper investigates whether transformers, in particular GPT-2, can learn to predict simple cellular automata, and also how they do it. This is an unexpected and quite interesting topic, and the investigation is well executed. The uncommon topic/perspective makes it worthy of... | {
"decision": "Accept (Poster)"
} |
Igm9bbkzHC | 2411.07404v1 | Controllable Context Sensitivity and the Knob Behind It | {
"content": "## Abstract\n\nAbstract When making predictions, a language model must trade off how much it relies on its context vs. its prior knowledge.\nChoosing how sensitive the model is to its context is a fundamental functionality, as it enables the model to excel at tasks like retrieval-augmented generation an... | [
{
"id": "EieDknyreI",
"initial_rating": 5,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper proposes to localize model behavior when models override their parametric knowledge with knowledge provided in context vs. when they do not override. T... | {
"rating": "3;5;6;8",
"rating_avg": 5.5,
"confidence": "2;4;3;3",
"confidence_avg": 3,
"soundness": "2;3;3;4",
"soundness_avg": 3,
"contribution": "1;2;3;4",
"contribution_avg": 2.5,
"presentation": "3;2;3;4",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.591130"
} | {
"id": "leej2duITo",
"metareview": "This paper identifies a linear subspace in a single layer that controls how strongly a model is sensitive to its context vs. prior knowledge. By modifying activations at this layer, the paper achieves a \"knob\" in a model as well as base models from the same family. The model ... | {
"decision": "Accept (Poster)"
} |
IgrLJslvxa | 2410.08811v1 | PoisonBench: Assessing Large Language Model Vulnerability to Data Poisoning | {
"content": "## Abstract\n\nAbstract Preference learning is a central component for aligning current LLMs, but this process can be vulnerable to data poisoning attacks. To address this concern, we introduce PoisonBench , a benchmark for evaluating large language models’ susceptibility to data poisoning during prefer... | [
{
"id": "3S5mZOXW9c",
"initial_rating": 3,
"confidence": 3,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "This paper presents POISONBENCH, a benchmark designed to assess the vulnerability of large language models (LLMs) to data poisoning attacks during preference lear... | {
"rating": "3;5;5;5;5;5",
"rating_avg": 4.666666666666667,
"confidence": "3;4;4;3;4;4",
"confidence_avg": 3.6666666666666665,
"soundness": "2;2;3;2;2;2",
"soundness_avg": 2.1666666666666665,
"contribution": "2;2;3;2;3;2",
"contribution_avg": 2.3333333333333335,
"presentation": "3;3;3;3;3;3",
"prese... | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.592221"
} | {
"id": "v9aqZBu827",
"metareview": "The paper introduces PoisonBench, a benchmark for evaluating LLMs' susceptibility to data poisoning during preference learning. \nThe authors addressed some concerns raised by the reviewers, such as the scalability and defense strategy.\nHowever, most reviewers remain concerned ... | {
"decision": "Reject"
} |
IiagjrJNwF | 2405.06394v2 | Memory Mosaics | {
"content": "## Abstract\n\nAbstract Memory Mosaics are networks of associative memories working in concert to achieve a prediction task of interest. Like transformers, memory mosaics possess compositional capabilities and in-context learning capabilities. Unlike transformers, memory mosaics achieve these capabiliti... | [
{
"id": "Ufe9bwzCj2",
"initial_rating": 3,
"confidence": 3,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "\"Memory Mosaics\" introduces an architecture where multiple associative memories are employed in unison to enhance predictive capabilities for tasks like languag... | {
"rating": "3;3;3;5",
"rating_avg": 3.5,
"confidence": "4;4;3;3",
"confidence_avg": 3.5,
"soundness": "2;2;2;3",
"soundness_avg": 2.25,
"contribution": "2;2;2;3",
"contribution_avg": 2.25,
"presentation": "2;1;2;3",
"presentation_avg": 2
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.593026"
} | {
"id": "QvUNN1EFXk",
"metareview": "This paper proposes a neural network architecture that composes of associative memory units. Compared to transformers, it achieves a new capability, called predictive disentanglement, under a meta-learning interpretation of the training process. The paper shows better interpreta... | {
"decision": "Accept (Poster)"
} |
IiwyThOFXL | 2406.10673v1 | SemanticMIM: Marring Masked Image Modeling with Semantics Compression for General Visual Representation | {
"content": "## Abstract\n\nAbstract This paper represents a neat yet effective framework, named SemanticMIM, to integrate the advantages of masked image modeling (MIM) and contrastive learning (CL) for general visual representation. We conduct a thorough comparative analysis between CL and MIM, revealing that their... | [
{
"id": "ttsuB3udZ5",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 2,
"summary": "This paper first compares the pros and cons of two related but different lines of work: masked image modeling (MIM) and contrastive learning (CL). They then propo... | {
"rating": "3;5;6;6",
"rating_avg": 5,
"confidence": "4;3;3;3",
"confidence_avg": 3.25,
"soundness": "3;2;3;3",
"soundness_avg": 2.75,
"contribution": "2;3;3;3",
"contribution_avg": 2.75,
"presentation": "1;2;3;2",
"presentation_avg": 2
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.594198"
} | {
"id": "CNHUDexpJA",
"metareview": "This paper presents a visual pre-training algorithm that unifies masked image modeling and contrastive learning. It first formulates these two types of pre-training methods using one framework, and then proposes a new pipeline that uses compressed image tokens to guide the model... | {
"decision": "Reject"
} |
IjVCcykKdr | 2312.02521v3 | RetriBooru: Leakage-Free Retrieval of Conditions from Reference Images for Subject-Driven Generation | {
"content": "## Abstract\n\nAbstract Diffusion-based methods have demonstrated remarkable capabilities in generating a diverse array of high-quality images, sparking interests for styled avatars, virtual try-on, and more. Previous methods use the same reference image as the target. An overlooked aspect is the leakag... | [
{
"id": "CJ9MzWRj9A",
"initial_rating": 5,
"confidence": 3,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "This paper proposes a large-scale dataset RetriBooru for training subject-preserving generation model, for each identity, it has multiple image include various sc... | {
"rating": "3;5;5;5",
"rating_avg": 4.5,
"confidence": "4;4;4;3",
"confidence_avg": 3.75,
"soundness": "2;3;2;2",
"soundness_avg": 2.25,
"contribution": "2;2;3;2",
"contribution_avg": 2.25,
"presentation": "2;2;3;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.595160"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
IjbXZdugdj | 2411.04165v1 | Bio-xLSTM: Generative modeling, representation and in-context learning of biological and chemical sequences | {
"content": "## Abstract\n\nAbstract Language models for biological and chemical sequences\nenable crucial applications such as drug discovery, protein engineering, and precision medicine.\nCurrently, these language models are predominantly based on Transformer architectures.\nWhile Transformers have yielded impress... | [
{
"id": "hRKvX432Ki",
"initial_rating": 8,
"confidence": 3,
"soundness": 4,
"contribution": 3,
"presentation": 4,
"summary": "The paper introduces Bio-xLSTM, a set of models that is tailored towards modeling of biological sequences. Specifically, the authors apply xLSTMs to different tas... | {
"rating": "3;6;6;8",
"rating_avg": 5.75,
"confidence": "5;3;2;3",
"confidence_avg": 3.25,
"soundness": "2;3;3;4",
"soundness_avg": 3,
"contribution": "1;3;2;3",
"contribution_avg": 2.25,
"presentation": "2;3;3;4",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.597291"
} | {
"id": "47yyhMwKGk",
"metareview": "This paper introduces Bio-xLSTM, a set of models leveraging the xLSTM architecture for modeling biological and chemical sequences across three domains: DNA, proteins, and small molecules. The architecture addresses limitations of Transformers in terms of quadratic scaling and me... | {
"decision": "Accept (Poster)"
} |
IjduZQK8gM | 2410.17174v1 | From Attention to Activation: Unraveling the Enigmas of Large Language Models | {
"content": "## Abstract\n\nAbstract We study two strange phenomena in auto-regressive Transformers:\n(1) the dominance of the first token in attention heads;\n(2) the occurrence of large outlier activations in the hidden states.\nWe find that popular large language models,\nsuch as Llama attend maximally to the fir... | [
{
"id": "h9JchJP2gr",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 1,
"presentation": 3,
"summary": "This paper studies two phenomena in transformer-based models. The first is the strong dominance of the first token in the attention maps, and the second is the pr... | {
"rating": "5;5;6",
"rating_avg": 5.333333333333333,
"confidence": "4;4;3",
"confidence_avg": 3.6666666666666665,
"soundness": "1;2;3",
"soundness_avg": 2,
"contribution": "4;1;3",
"contribution_avg": 2.6666666666666665,
"presentation": "3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.598299"
} | {
"id": "60TNNOGnyw",
"metareview": "The paper offers two main contributions the dominance of first tokens in the self-attention maps, and the presence of outlier dimensions throughout the hidden representations of these models.\n\nStrengths:\nAddresses known issues in the literature, offers practical techniques\nC... | {
"decision": "Accept (Poster)"
} |
IkZVDzdC8M | 2411.07231v1 | Watermark Anything With Localized Messages | {
"content": "## Abstract\n\nAbstract Image watermarking methods are not tailored to handle small watermarked areas.\nThis restricts applications in real-world scenarios where parts of the image may come from different sources or have been edited.\nWe introduce a deep-learning model for localized image watermarking, ... | [
{
"id": "SoQOLNFZRz",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper proposes a new method, Watermark Anything Model (WAM), for adding watermarks to images. The method demonstrates state-of-the-art performance even under... | {
"rating": "5;5;5;5;6;6",
"rating_avg": 5.333333333333333,
"confidence": "4;3;4;4;4;3",
"confidence_avg": 3.6666666666666665,
"soundness": "3;3;2;2;3;3",
"soundness_avg": 2.6666666666666665,
"contribution": "2;2;2;2;3;3",
"contribution_avg": 2.3333333333333335,
"presentation": "2;3;3;2;2;3",
"prese... | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.599286"
} | {
"id": "9ypWZ5mEzK",
"metareview": "This paper works on watermarking. Authors proposed a deep-learning model for localized image watermarking, WAM, which could extract different watermarks from different regions. \n\nThis paper was reviewed by 6 reviewers and got mixed scores as three 8, two 5, one 3. \n\nStrength... | {
"decision": "Accept (Poster)"
} |
ImpeMDJfVL | 2410.04171v2 | IV-mixed Sampler: Leveraging Image Diffusion Models for Enhanced Video Synthesis | {
"content": "## Abstract\n\nAbstract The multi-step sampling mechanism, a key feature of visual diffusion models, has significant potential to replicate the success of OpenAI’s Strawberry in enhancing performance by increasing the inference computational cost. Sufficient prior studies have demonstrated that correctl... | [
{
"id": "7lLaTJE2pw",
"initial_rating": 6,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "The paper presents a novel sampling scheme for video diffusion models that leverages pre-train image diffusion models, as they generally have higher visual fideli... | {
"rating": "5;5;6",
"rating_avg": 5.333333333333333,
"confidence": "3;3;4",
"confidence_avg": 3.3333333333333335,
"soundness": "2;3;3",
"soundness_avg": 2.6666666666666665,
"contribution": "2;2;3",
"contribution_avg": 2.3333333333333335,
"presentation": "3;1;3",
"presentation_avg": 2.33333333333333... | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.600101"
} | {
"id": "ExhKd7aXgv",
"metareview": "The paper proposes IV-mixed Sampler, a method for improving the quality and capabilities of video diffusion models by integrating an image diffusion model in the sampling process (since image models are known to have higher visual quality)---effectively mixing the spatial/visual... | {
"decision": "Accept (Poster)"
} |
InWaCoIMMN | 2303.00333v4 | Competence-Based Analysis of Language Models | {
"content": "## Abstract\n\nAbstract Despite the recent successes of large, pretrained neural language models (LLMs), comparatively little is known about the representations of linguistic structure they learn during pretraining, which can lead to unexpected behaviors in response to prompt variation or distribution s... | [
{
"id": "jfXFsb5g9y",
"initial_rating": 3,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "This paper proposes analyzing LLM capabilities via competence-based testing, specifically via intervening on the model’s representations via gradient-based advers... | {
"rating": "3;3;3;3",
"rating_avg": 3,
"confidence": "3;4;4;3",
"confidence_avg": 3.5,
"soundness": "2;3;2;3",
"soundness_avg": 2.5,
"contribution": "1;2;2;2",
"contribution_avg": 1.75,
"presentation": "2;2;2;3",
"presentation_avg": 2.25
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.600924"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
Io9yFt7XH7 | 2409.00101v1 | NeuroLM: A Universal Multi-task Foundation Model for Bridging the Gap between Language and EEG Signals | {
"content": "## Abstract\n\nAbstract Recent advancements for large-scale pre-training with neural signals such as electroencephalogram (EEG) have shown promising results, significantly boosting the development of brain-computer interfaces (BCIs) and healthcare. However, these pre-trained models often require full fi... | [
{
"id": "c2zTdVmlTZ",
"initial_rating": 8,
"confidence": 4,
"soundness": 4,
"contribution": 3,
"presentation": 4,
"summary": "The paper presents NeuroLM, an innovative multi-task foundation model that bridges the gap between electroencephalogram (EEG) signals and language models. By conv... | {
"rating": "5;5;6;8",
"rating_avg": 6,
"confidence": "3;3;4;4",
"confidence_avg": 3.5,
"soundness": "2;3;3;4",
"soundness_avg": 3,
"contribution": "2;3;3;3",
"contribution_avg": 2.75,
"presentation": "2;3;2;4",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.602461"
} | {
"id": "46Wbq2TYsg",
"metareview": "The paper presents an innovative method for predicting on diverse EEG tasks within a single unified model without the need for task-specific fine-tuning.\n\nPros:\n- The approach is considered original and novel\n- Models are trained on a very large EEG corpus\n- Results are pro... | {
"decision": "Accept (Poster)"
} |
Ip6UwB35uT | 2409.16829v1 | Conditional Testing based on Localized Conformal $p$-values | {
"content": "## Abstract\n\nAbstract In this paper, we address conditional testing problems through the conformal inference framework. We define the localized conformal p 𝑝 p italic_p -values by inverting prediction intervals and prove their theoretical properties. These defined p 𝑝 p italic_p -values are then app... | [
{
"id": "d0R0h8Rb8n",
"initial_rating": 6,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "The paper introduces localized conformal $p$-values and applies them to several conditional testing problems, including conditional outlier detection, conditional... | {
"rating": "5;6;6;6",
"rating_avg": 5.75,
"confidence": "4;3;3;4",
"confidence_avg": 3.5,
"soundness": "3;3;3;3",
"soundness_avg": 3,
"contribution": "2;3;2;3",
"contribution_avg": 2.5,
"presentation": "2;3;3;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.603337"
} | {
"id": "YaGoyp4whd",
"metareview": "This paper applies conformal inference techniques to perform conditional testing. They define a new localized conformal p value, and use this to perform several practical conditional tests including outlier detecting and label screening. Reviewers found this to be novel and some... | {
"decision": "Accept (Poster)"
} |
Ir6JxcuP6H | 2410.23287v1 | ReferEverything: Towards segmenting everything we can speak of in videos | {
"content": "## Abstract\n\nAbstract We present REM, a framework for segmenting a wide range of concepts in video that can be described through natural language. Our method capitalizes on visual-language representations learned by video diffusion models on Internet-scale datasets. A key insight of our approach is pr... | [
{
"id": "S9gi1tcvxs",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "The REM paper introduces a novel framework for Referral Video Segmentation (RVS) that uses a video-text diffusion model to generalize segmentation to a wide range... | {
"rating": "3;5;5;6;6",
"rating_avg": 5,
"confidence": "5;4;3;4;3",
"confidence_avg": 3.8,
"soundness": "3;3;2;3;3",
"soundness_avg": 2.8,
"contribution": "2;2;2;2;3",
"contribution_avg": 2.2,
"presentation": "2;3;3;3;3",
"presentation_avg": 2.8
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.604757"
} | {
"id": "KIta9iuouC",
"metareview": "This paper receives mixed reviews from five reviewers.\n\nThe positive reviewers think the new setting is interesting, and it evaluates models on dynamic processes.\n\nThe negative reviewers were concerned about the novelty of technical parts, ablation study details, and several... | {
"decision": "Reject"
} |
Iu8RytBaji | 2409.17407v1 | Post-hoc Reward Calibration: A Case Study on Length Bias | {
"content": "## Abstract\n\nAbstract Reinforcement Learning from Human Feedback aligns the outputs of Large Language Models with human values and preferences. Central to this process is the reward model (RM), which translates human feedback into training signals for optimising LLM behaviour.\nHowever, RMs can develo... | [
{
"id": "spv9Tz5TUI",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "This paper proposes a novel method, “Post-hoc Reward Calibration,” to mitigate biases in reward models (RMs), particularly focusing on LLM output text length bias... | {
"rating": "5;5;6;8",
"rating_avg": 6,
"confidence": "4;4;4;3",
"confidence_avg": 3.75,
"soundness": "2;2;3;4",
"soundness_avg": 2.75,
"contribution": "2;2;2;3",
"contribution_avg": 2.25,
"presentation": "3;3;2;4",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.605661"
} | {
"id": "rDPAAvBZPs",
"metareview": "This paper studies the length bias in reward modeling and RLHF, which is a well-known issue in learned reward models and language model preferences. They provide a method of \"Post-hoc Reward Calibration\" to address it and obtained promising empirical results.\n\nStrengths: \n... | {
"decision": "Accept (Poster)"
} |
IwPXYk6BV9 | 2405.15150v1 | Enhancing Learning with Label Differential Privacy by Vector Approximation | {
"content": "## Abstract\n\nAbstract Label differential privacy (DP) is a framework that protects the privacy of labels in training datasets, while the feature vectors are public. Existing approaches protect the privacy of labels by flipping them randomly, and then train a model to make the output approximate the pr... | [
{
"id": "WHHpNq8ew2",
"initial_rating": 8,
"confidence": 4,
"soundness": 4,
"contribution": 3,
"presentation": 3,
"summary": "This paper consider the problem of supervised multi-class learning problem under the setting of \"label differential privacy\" (LDP). This is a setting that is we... | {
"rating": "3;6;6;8",
"rating_avg": 5.75,
"confidence": "3;4;4;4",
"confidence_avg": 3.75,
"soundness": "2;3;3;4",
"soundness_avg": 3,
"contribution": "1;3;3;3",
"contribution_avg": 2.5,
"presentation": "3;2;4;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Spotlight",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.606777"
} | {
"id": "jJDXiz8cdr",
"metareview": "## Summary of Contributions\n\nThis paper studies label differential privacy (label-DP) setting, where we want to train an ML model and the examples are public but the labels are sensitive and should be protected with differential privacy. The paper proposes a simple but effecti... | {
"decision": "Accept (Spotlight)"
} |
IwgmgidYPS | 2408.02900v1 | MedTrinity-25M: A Large-scale Multimodal Dataset with Multigranular Annotations for Medicine | {
"content": "## Abstract\n\nAbstract This paper introduces MedTrinity-25M, a comprehensive, large-scale multimodal dataset for medicine, covering over 25 million images across 10 modalities, with multigranular annotations for more than 65 diseases. These enriched annotations encompass both global textual information... | [
{
"id": "uMr3NwKUyt",
"initial_rating": 6,
"confidence": 5,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "This paper introduces MedTrinity-25M, a comprehensive and large-scale multimodal dataset for medicine, covering over 25 million images across 10 modalities with d... | {
"rating": "5;5;6;6",
"rating_avg": 5.5,
"confidence": "5;4;4;5",
"confidence_avg": 4.5,
"soundness": "3;3;3;2",
"soundness_avg": 2.75,
"contribution": "2;2;3;2",
"contribution_avg": 2.25,
"presentation": "2;3;3;2",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.607973"
} | {
"id": "JNZoNgziY4",
"metareview": "The paper introduces MedTrinity-25M, a large-scale multimodal medical dataset comprising over 25 million images across 10 modalities, with multigranular annotations covering more than 65 diseases. The annotations include both global information, such as modality and organ detect... | {
"decision": "Accept (Poster)"
} |
Iyrtb9EJBp | 2409.11242v2 | Measuring and Enhancing Trustworthiness of LLMs in RAG through Grounded Attributions and Learning to Refuse | {
"content": "## Abstract\n\nAbstract LLMs are an integral component of retrieval-augmented generation (RAG) systems. While many studies focus on evaluating the overall quality of end-to-end RAG systems, there is a gap in understanding the appropriateness of LLMs for the RAG task. To address this, we introduce Trust-... | [
{
"id": "W4qmKebIv1",
"initial_rating": 8,
"confidence": 4,
"soundness": 4,
"contribution": 4,
"presentation": 4,
"summary": "Authors present a study of ‘grounded’ RAG in LLMs, i.e., a method for evaluating and aligning LLMs for RAG to increase the correctness of RAG, that LLMs cite the ... | {
"rating": "6;8;8;8",
"rating_avg": 7.5,
"confidence": "3;3;2;4",
"confidence_avg": 3,
"soundness": "2;3;3;4",
"soundness_avg": 3,
"contribution": "3;3;4;4",
"contribution_avg": 3.5,
"presentation": "3;3;3;4",
"presentation_avg": 3.25
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Oral",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.609604"
} | {
"id": "iHNOcvSDdm",
"metareview": "This paper introduces a method (Trust-Align) for enhancing a model’s response across three dimensions: trustworthiness, citation groundedness, and refusal groundedness, measured by their groundedness metric, Trust-Score. The authors evaluate the proposed metric and alignment me... | {
"decision": "Accept (Oral)"
} |
Iz75SDbRmm | 2409.08202v1 | What Makes a Maze Look Like a Maze? | {
"content": "## Abstract\n\nAbstract A unique aspect of human visual understanding is the ability to flexibly interpret abstract concepts: acquiring lifted rules explaining what they symbolize, grounding them across familiar and unfamiliar contexts, and making predictions or reasoning about them. While off-the-shelf... | [
{
"id": "yPHTPoY1Va",
"initial_rating": 5,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "The paper introduces a framework called Deep Schema Grounding (DSG) and a VQA benchmark for evaluating abstract visual reasoning of VLMs. For a given image and a ... | {
"rating": "5;5;6;8;8",
"rating_avg": 6.4,
"confidence": "3;4;3;3;4",
"confidence_avg": 3.4,
"soundness": "3;3;3;3;4",
"soundness_avg": 3.2,
"contribution": "2;3;2;3;3",
"contribution_avg": 2.6,
"presentation": "3;3;3;3;4",
"presentation_avg": 3.2
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.610938"
} | {
"id": "Y8KPF6oJga",
"metareview": "This paper introduces Deep Schema Grounding (DSG), which enhances VLMs in abstract visual reasoning by using an LLM to generate schemas that decompose abstract concepts into smaller components, which are then grounded onto images for answering related questions. Experimental res... | {
"decision": "Accept (Poster)"
} |
IzQB2pIa3F | 2410.01731v1 | ComfyGen: Prompt-Adaptive Workflows for Text-to-Image Generation | {
"content": "## Abstract\n\nAbstract The practical use of text-to-image generation has evolved from simple, monolithic models to complex workflows that combine multiple specialized components. While workflow-based approaches can lead to improved image quality, crafting effective workflows requires significant expert... | [
{
"id": "1jHPUo09yH",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 3,
"presentation": 2,
"summary": "This paper introduces a novel approach to text-to-image generation by developing adaptive workflows that cater to specific user prompts, enhancing the quality of ... | {
"rating": "3;5;5;5;6",
"rating_avg": 4.8,
"confidence": "5;4;3;4;3",
"confidence_avg": 3.8,
"soundness": "3;3;2;2;3",
"soundness_avg": 2.6,
"contribution": "3;3;2;3;3",
"contribution_avg": 2.8,
"presentation": "3;4;3;2;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.612369"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
IzYczpPqKq | 2407.10207v2 | Learning to Steer Markovian Agents under Model Uncertainty | {
"content": "## Abstract\n\nAbstract Designing incentives for an adapting population is a ubiquitous problem in a wide array of economic applications and beyond. In this work, we study how to design additional rewards to steer multi-agent systems towards desired policies without prior knowledge of the agents’ underl... | [
{
"id": "Gfl2NcjNv5",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "This paper presents a framework for steering the behaviors of agents towards desired outcomes within a reinforcement learning setting, where the exact learning dy... | {
"rating": "3;5;8",
"rating_avg": 5.333333333333333,
"confidence": "3;3;4",
"confidence_avg": 3.3333333333333335,
"soundness": "3;3;3",
"soundness_avg": 3,
"contribution": "2;2;3",
"contribution_avg": 2.3333333333333335,
"presentation": "1;3;3",
"presentation_avg": 2.3333333333333335
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.615137"
} | {
"id": "k0xDShHSei",
"metareview": "This paper presents a framework for steering Markovian agents in multi-agent systems using reinforcement learning under model uncertainty, with strong theoretical guarantees and empirical validations. Despite some concerns about scalability to more complex and realistic settings... | {
"decision": "Accept (Poster)"
} |
J2Jyp1SZ0n | 2409.12959v1 | MMSearch: Benchmarking the Potential of Large Models as Multi-modal Search Engines | {
"content": "## Abstract\n\nAbstract The advent of Large Language Models (LLMs) has paved the way for AI search engines, e.g., SearchGPT, showcasing a new paradigm in human-internet interaction.\nHowever, most current AI search engines are limited to text-only settings, neglecting the multimodal user queries and the... | [
{
"id": "3IVoWL9Gas",
"initial_rating": 6,
"confidence": 4,
"soundness": 4,
"contribution": 4,
"presentation": 4,
"summary": "This work describes a new pipeline to empower LMMs with multimodal search capabilities called MMSearch-Engine, and introduce MMSearch, a comprehensive evaluation ... | {
"rating": "6;6;6;8",
"rating_avg": 6.5,
"confidence": "3;5;4;4",
"confidence_avg": 4,
"soundness": "3;3;4;3",
"soundness_avg": 3.25,
"contribution": "2;3;4;3",
"contribution_avg": 3,
"presentation": "3;3;4;4",
"presentation_avg": 3.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.616623"
} | {
"id": "3AyA5yeKNQ",
"metareview": "**Summary:**\n\nThis paper proposes a straightforward approach to developing a multi-modal search engine by integrating LLMs, Google Lens, and other tools. The core components of the system are re-querying, reranking, and summarization. The scope of the paper is well-defined and... | {
"decision": "Accept (Poster)"
} |
J2we1sVd9m | 2410.07617v1 | Prototype-based Optimal Transport for Out-of-Distribution Detection | {
"content": "## Abstract\n\nAbstract Detecting Out-of-Distribution (OOD) inputs is crucial for improving the reliability of deep neural networks in the real-world deployment. In this paper, inspired by the inherent distribution shift between ID and OOD data, we propose a novel method that leverages optimal transport... | [
{
"id": "gtXINTrsLG",
"initial_rating": 3,
"confidence": 5,
"soundness": 2,
"contribution": 3,
"presentation": 2,
"summary": "Summary:This paper proposes a new prototype-based optimal transfer (POT) method to detect out-of-distribution (OOD) data by measuring the optimal total transfer c... | {
"rating": "3;3;3;6;6",
"rating_avg": 4.2,
"confidence": "4;5;5;3;4",
"confidence_avg": 4.2,
"soundness": "2;3;2;3;4",
"soundness_avg": 2.8,
"contribution": "2;2;3;3;4",
"contribution_avg": 2.8,
"presentation": "2;2;2;3;3",
"presentation_avg": 2.4
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.617364"
} | {
"id": "52X8Ov29ny",
"metareview": "This submission focuses on out-of-distribution (OOD) detection and proposes a novel method based on prototypes and optimal transport. Specifically, the method first calculates the optimal transport distance from each test batch to the in-distribution (ID) prototypes. To improve ... | {
"decision": "Reject"
} |
J3xRByRqOz | 2410.22954v1 | RETRIEVAL-AUGMENTED GENERATION WITH ESTIMATION OF SOURCE RELIABILITY | {
"content": "## Abstract\n\nAbstract Retrieval-augmented generation (RAG) addresses key limitations of large language models (LLMs), such as hallucinations and outdated knowledge, by incorporating external databases. These databases typically consult multiple sources to encompass up-to-date and various information. ... | [
{
"id": "dU8AXdacmT",
"initial_rating": 3,
"confidence": 3,
"soundness": 2,
"contribution": 3,
"presentation": 4,
"summary": "These authors aim to address the challenge of heterogeneous source reliability in RAG, and propose a novel framework, Reliability-Aware RAG (RA-RAG), which incorp... | {
"rating": "3;3;5;5",
"rating_avg": 4,
"confidence": "5;3;3;4",
"confidence_avg": 3.75,
"soundness": "2;2;3;3",
"soundness_avg": 2.5,
"contribution": "2;3;1;3",
"contribution_avg": 2.25,
"presentation": "2;4;4;2",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.618106"
} | {
"id": "0AFMReQm3N",
"metareview": "This paper introudces a new RAG framework, named reliability-aware RAG (RA-RAG), which estimates the reliability of multiple sources and incorporates this information into both the retrieval and aggregation processes. Moreover, the authors also introduce a benchmark dataset desi... | {
"decision": "Reject"
} |
J5sUOvlLbQ | 2410.17146v1 | LiNeS: Post-training Layer Scaling Prevents Forgetting and Enhances Model Merging | {
"content": "## Abstract\n\nAbstract Large pre-trained models exhibit impressive zero-shot performance across diverse tasks, but fine-tuning often leads to catastrophic forgetting, where improvements on a target domain degrade generalization on other tasks. To address this challenge, we introduce LiNeS, Layer-increa... | [
{
"id": "pxTZwsbP4I",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 2,
"summary": "This paper introduces a post-training editing technique, LiNeS, designed to address catastrophic forgetting and facilitate model merging after fine-tuning. LiNeS ... | {
"rating": "5;5;6;6",
"rating_avg": 5.5,
"confidence": "4;4;3;3",
"confidence_avg": 3.5,
"soundness": "3;3;3;3",
"soundness_avg": 3,
"contribution": "2;3;3;3",
"contribution_avg": 2.75,
"presentation": "3;3;4;2",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.618920"
} | {
"id": "It5ZH54otz",
"metareview": "Summary: LiNeS is a post-training method that applies layer-wise scaling of weight distances, preserving general features while enhancing task-specific deeper features in single- or multi-task settings. It can mitigate catastrophic forgetting and improve model merging.\n\nStreng... | {
"decision": "Accept (Poster)"
} |
J6Xgra2bE5 | 2410.17222v1 | Context-aware Prompt Tuning: Advancing In-Context Learning with Adversarial Methods | {
"content": "## Abstract\n\nAbstract Fine-tuning Large Language Models (LLMs) typically involves updating at least a few billions of parameters.\nA more parameter-efficient approach is Prompt Tuning (PT), which updates only a few learnable tokens, and differently, In-Context Learning (ICL) adapts the model to a new ... | [
{
"id": "MgqFb7jOt5",
"initial_rating": 3,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "This paper proposes Context-aware Prompt Tuning (CPT) as a new method to address the overfitting problem of prompt tuning. The optimization of CPT consists of a n... | {
"rating": "3;3;5;6",
"rating_avg": 4.25,
"confidence": "4;4;4;4",
"confidence_avg": 4,
"soundness": "2;2;2;3",
"soundness_avg": 2.25,
"contribution": "2;2;2;3",
"contribution_avg": 2.25,
"presentation": "2;2;2;3",
"presentation_avg": 2.25
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.619719"
} | {
"id": "vMa9swnfXW",
"metareview": "This paper introduces Context-aware Prompt Tuning (CPT), a method combining In-Context Learning (ICL), Prompt Tuning (PT), and adversarial optimization to address overfitting in few-shot learning tasks. The approach is validated on multiple classification datasets and different ... | {
"decision": "Reject"
} |
J8YWCBPgx7 | 2410.04844v1 | PostEdit: Posterior Sampling for Efficient Zero-Shot Image Editing | {
"content": "## Abstract\n\nAbstract In the field of image editing, three core challenges persist: controllability, background preservation, and efficiency. Inversion-based methods rely on time-consuming optimization to preserve the features of the initial images, which results in low efficiency due to the requireme... | [
{
"id": "9eeEF84DpB",
"initial_rating": 6,
"confidence": 5,
"soundness": 3,
"contribution": 3,
"presentation": 2,
"summary": "This paper tackles the issues caused by the unconditional term in Classifier-Free Guidance (CFG) by integrating the theory of posterior sampling to improve recons... | {
"rating": "5;6;6;6",
"rating_avg": 5.75,
"confidence": "4;5;4;5",
"confidence_avg": 4.5,
"soundness": "3;3;3;3",
"soundness_avg": 3,
"contribution": "3;3;3;3",
"contribution_avg": 3,
"presentation": "3;3;3;2",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.620418"
} | {
"id": "w4rDVZEFh9",
"metareview": "This paper introduces PostEdit, a novel method for zero-shot image editing that leverages posterior sampling to achieve both efficiency and background preservation. The paper claims that PostEdit overcomes the limitations of existing inversion-based and inversion-free methods by... | {
"decision": "Accept (Poster)"
} |
J8yH8ontdq | 2410.03335v1 | Audio-Agent: Leveraging LLMs For Audio Generation, Editing and Composition | {
"content": "## Abstract\n\nAbstract We introduce Audio-Agent, a multimodal framework for audio generation, editing and composition based on text or video inputs. Conventional approaches for text-to-audio (TTA) tasks often make single-pass inferences from text descriptions. While straightforward, this design struggl... | [
{
"id": "D7v5wjZxP2",
"initial_rating": 1,
"confidence": 5,
"soundness": 2,
"contribution": 1,
"presentation": 2,
"summary": "This work introduces Audio-Agent, an LLM-based framework for audio content creation, comprising two main components. For text-to-audio (TTA) tasks, Audio-Agent le... | {
"rating": "1;3;5;8",
"rating_avg": 4.25,
"confidence": "5;4;4;4",
"confidence_avg": 4.25,
"soundness": "2;2;3;4",
"soundness_avg": 2.75,
"contribution": "1;2;2;3",
"contribution_avg": 2,
"presentation": "2;2;3;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.621237"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
J9Ofr1PmvX | 2410.17050v1 | UnSTAR: Unlearning with Self-Taught Anti-Sample Reasoning for LLMs | {
"content": "## Abstract\n\nAbstract The key components of machine learning are data samples for training, model for learning patterns, and loss function for optimizing accuracy. Analogously, unlearning can potentially be achieved through anti-data-samples (or anti-samples), unlearning method, and reversed loss func... | [
{
"id": "Q0CH86Lh0d",
"initial_rating": 8,
"confidence": 2,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "The paper focuses on the idea to use anti-samples coupled with reasoning tuning to enable target machine unlearning. Given a question, the methods utilizes langua... | {
"rating": "3;5;5;5;5;6",
"rating_avg": 4.833333333333333,
"confidence": "3;3;4;4;3;3",
"confidence_avg": 3.3333333333333335,
"soundness": "2;2;2;3;3;3",
"soundness_avg": 2.5,
"contribution": "2;2;2;3;3;3",
"contribution_avg": 2.5,
"presentation": "2;3;2;3;3;3",
"presentation_avg": 2.66666666666666... | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.621872"
} | {
"id": "illDiG8Wch",
"metareview": "This submission falls slightly below the acceptance threshold, as reflected in the final review ratings: 8 (Reviewer DtLx, with lower confidence 2), 5 (Reviewer fwuc), 5 (Reviewer qhwt), 5 (Reviewer 9vHW), 5 (Reviewer 84co), and 5 (Reviewer F1yx). While the authors' rebuttal add... | {
"decision": "Reject"
} |
J9VogDTa1W | 2410.13054v1 | Systems with Switching Causal Relations: A Meta-Causal Perspective | {
"content": "## Abstract\n\nAbstract Most work on causality in machine learning assumes that causal relationships are driven by a constant underlying process. However, the flexibility of agents’ actions or tipping points in the environmental process can change the qualitative dynamics of the system. As a result, new... | [
{
"id": "77OiHD9agn",
"initial_rating": 8,
"confidence": 3,
"soundness": 4,
"contribution": 3,
"presentation": 3,
"summary": "The paper introduces the concept of *meta-causal states*, which attempt to represent how the underlying dynamics of a causal model can change based on context. Th... | {
"rating": "3;6;6;8",
"rating_avg": 5.75,
"confidence": "3;2;3;3",
"confidence_avg": 2.75,
"soundness": "2;3;4;4",
"soundness_avg": 3.25,
"contribution": "3;4;2;3",
"contribution_avg": 3,
"presentation": "2;3;3;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Spotlight",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.622674"
} | {
"id": "cGjHOHYVbs",
"metareview": "This paper proposes meta-causal models (MCM) to handle changing causal structures, extending classical causal models so that the system’s “causal equations” can shift based on state or context. Reviewers praised the novelty of modeling how causal mechanisms themselves emerge and... | {
"decision": "Accept (Spotlight)"
} |
J9eKm7j6KD | 2406.11624v2 | Words in Motion: Extracting Interpretable Control Vectors for Motion Transformers | {
"content": "## Abstract\n\nAbstract Motion forecasting transforms sequences of past movements and environment context into future motion.\nRecent methods rely on learned representations, resulting in hidden states that are difficult to interpret.\nIn this work, we use natural language to quantize motion features in... | [
{
"id": "rwUqBHskmW",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "The paper focuses on the interpretability and control of transformer-based motion forecasting models. The authors aim to interpret hidden states of motion transfo... | {
"rating": "3;3;5;5;6",
"rating_avg": 4.4,
"confidence": "3;5;4;4;3",
"confidence_avg": 3.8,
"soundness": "2;2;2;2;3",
"soundness_avg": 2.2,
"contribution": "2;2;2;2;3",
"contribution_avg": 2.2,
"presentation": "2;1;2;3;3",
"presentation_avg": 2.2
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.623342"
} | {
"id": "pysTq2wdtk",
"metareview": "The problem studied in this submission is considered interesting by the majority of reviewers, attracting favourable remarks about clear motivation and broader interest in studying interpretability techniques for transformers beyond the language domain. In addition, other positi... | {
"decision": "Accept (Poster)"
} |
JCCPtPDido | 2410.06024v1 | Jet Expansions of Residual Computation | {
"content": "## Abstract\n\nAbstract We introduce a framework for expanding residual computational graphs using jets , operators that generalize truncated Taylor series.\nOur method provides a systematic approach to disentangle contributions of different computational paths to model predictions.\nIn contrast to exis... | [
{
"id": "qH3R4MIlHX",
"initial_rating": 5,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 4,
"summary": "The paper introduces a method for expanding residual networks, such as Transformers, using jets—operators that generalize truncated Taylor series. This approach a... | {
"rating": "3;5;6;6",
"rating_avg": 5,
"confidence": "3;4;2;3",
"confidence_avg": 3,
"soundness": "2;4;3;3",
"soundness_avg": 3,
"contribution": "3;3;3;2",
"contribution_avg": 2.75,
"presentation": "1;2;3;4",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.624100"
} | {
"id": "zu5oeZflCH",
"metareview": "In this work, the authors introduce a framework for expanding residual networks into linear residual networks using Taylor-type expansions (jets). These jets encode path information through the computational graph which can then be interpreted to better understand higher-order b... | {
"decision": "Reject"
} |
JCiF03qnmi | 2410.21676v2 | How Does Critical Batch Size Scale in Pre-training? | {
"content": "## Abstract\n\nAbstract Training large-scale models under given resources requires careful design of parallelism strategies.\nIn particular, the efficiency notion of critical batch size (CBS), concerning the compromise between time and compute, marks the threshold beyond which greater data parallelism l... | [
{
"id": "9o8TV7sjg0",
"initial_rating": 8,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "The paper investigates the scaling behavior of critical batch size in the pre-training of autoregressive language models.\nThey first define the critical batch si... | {
"rating": "3;5;5;8;8",
"rating_avg": 5.8,
"confidence": "4;4;3;4;3",
"confidence_avg": 3.6,
"soundness": "1;3;3;4;3",
"soundness_avg": 2.8,
"contribution": "3;3;2;4;3",
"contribution_avg": 3,
"presentation": "3;3;3;4;3",
"presentation_avg": 3.2
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.625305"
} | {
"id": "XBNnrH6caq",
"metareview": "The paper explores the scaling behavior of critical batch size (CBS) in autoregressive language model pretraining. CBS is shown to scale primarily with dataset size rather than model size, a finding supported by theoretical analysis. Empirical studies span models from 85M to 1.2... | {
"decision": "Accept (Poster)"
} |
JDzTI9rKls | 2403.04453v2 | Efficient Off-Policy Learning for High-Dimensional Action Spaces | {
"content": "## Abstract\n\nAbstract Existing off-policy reinforcement learning algorithms often rely on an explicit state-action-value function representation, which can be problematic in high-dimensional action spaces due to the curse of dimensionality.\nThis reliance results in data inefficiency as maintaining a ... | [
{
"id": "qaRC6AU2Ix",
"initial_rating": 8,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 4,
"summary": "This work proposes a fully off-policy reinforcement learning algorithm based on policy gradient and importance sampling. Crucially, the algorithm does not attempt... | {
"rating": "3;5;8;8",
"rating_avg": 6,
"confidence": "4;4;4;3",
"confidence_avg": 3.75,
"soundness": "3;3;3;3",
"soundness_avg": 3,
"contribution": "2;2;3;2",
"contribution_avg": 2.25,
"presentation": "2;3;3;4",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.627515"
} | {
"id": "H1iuB8Svcl",
"metareview": "The paper proposes a new technique for offline RL that learns a state value function instead of state-action Q-function. Combined with importance sampling, the new technique is a policy gradient technique that can leverage off-policy data. The strength of this work is the simp... | {
"decision": "Accept (Poster)"
} |
JEehcb48Vp | 2408.16326v2 | Critic-CoT: Boosting the reasoning abilities of large language model via Chain-of-Thought Critic | {
"content": "## Abstract\n\nAbstract Self-critic has become a crucial mechanism for enhancing the reasoning performance of LLMs.\nHowever, current approaches mainly involve basic prompts for intuitive instance-level feedback, which resembles System-1 processes and limits the reasoning capabilities. Moreover, there i... | [
{
"id": "ytT5y9cxBQ",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "This paper proposes a CoT-based critic framework, including training data collections and two different inference methods (reflection or filter).",
"strengths... | {
"rating": "5;5;5;5",
"rating_avg": 5,
"confidence": "4;4;3;3",
"confidence_avg": 3.5,
"soundness": "3;3;2;3",
"soundness_avg": 2.75,
"contribution": "3;2;1;2",
"contribution_avg": 2,
"presentation": "2;3;2;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.628443"
} | {
"id": "EJ2lZolsdc",
"metareview": "This paper proposes Critic-CoT, a method designed to improve LLMs' reasoning by developing a stepwise, chain-of-thought (CoT) critique process. The approach improves LLMs to critique and refine their solutions automatically. Specifically, the authors generate step-wise feedback ... | {
"decision": "Reject"
} |
JEmNgjuQHU | 2407.05986v1 | KidSat: satellite imagery to map childhood poverty | {
"content": "## Abstract\n\nAbstract Satellite imagery has emerged as an important tool to analyse demographic, health, and development indicators. While various deep learning models have been built for these tasks, each is specific to a particular problem, with few standard benchmarks available. We propose a new da... | [
{
"id": "uDPQM9PoR7",
"initial_rating": 3,
"confidence": 5,
"soundness": 2,
"contribution": 1,
"presentation": 1,
"summary": "The paper \"KidSat: Satellite Imagery to Map Childhood Poverty\" presents a dataset and benchmarking framework aimed at improving the analysis of child poverty in... | {
"rating": "1;1;3;3",
"rating_avg": 2,
"confidence": "5;4;4;5",
"confidence_avg": 4.5,
"soundness": "2;2;2;2",
"soundness_avg": 2,
"contribution": "1;1;3;1",
"contribution_avg": 1.5,
"presentation": "3;2;2;1",
"presentation_avg": 2
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.631859"
} | {
"id": "L2RJaBVDeD",
"metareview": "Dear authors,\n\nThank you for submitting the draft. The reviewers' rankings indicated that the draft is not ready for publication at this stage.\n\nThe draft proposes a new dataset to predict child poverty from satellite imagery. Results are presented for both locations outsid... | {
"decision": "Reject"
} |
JFk8F7w8Iz | 2405.13541v1 | Annotation-Efficient Language Model Alignment via Diverse and Representative Response Texts | {
"content": "## Abstract\n\nAbstract Preference optimization is a standard approach to fine-tuning large language models to align with human preferences.\nThe quality, diversity, and quantity of the preference dataset are critical to the effectiveness of preference optimization.\nHowever, obtaining a large amount of... | [
{
"id": "aI7RWEJLnd",
"initial_rating": 3,
"confidence": 5,
"soundness": 1,
"contribution": 1,
"presentation": 2,
"summary": "The paper proposes to select generations for preference annotation by optimizing an objective that combines representativeness (in the sense of the selected gener... | {
"rating": "3;3;5;8",
"rating_avg": 4.75,
"confidence": "4;5;4;4",
"confidence_avg": 4.25,
"soundness": "2;1;2;4",
"soundness_avg": 2.25,
"contribution": "2;1;2;3",
"contribution_avg": 2,
"presentation": "3;2;3;4",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.632680"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
JIGuWpQcqO | 2410.08063v1 | Reversible Decoupling Network for Single Image Reflection Removal | {
"content": "## Abstract\n\nAbstract Recent deep-learning-based approaches to single-image reflection removal have shown promising advances, primarily for two reasons: 1) the utilization of recognition-pretrained features as inputs, and 2) the design of dual-stream interaction networks. However, according to the Inf... | [
{
"id": "ZqtPRpQpDO",
"initial_rating": 3,
"confidence": 5,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "This work addresses the problem of single-image reflection removal, introducing a novel Reversible Decoupling Network (RDNet) as its core approach. The RDNet is b... | {
"rating": "3;3;5",
"rating_avg": 3.6666666666666665,
"confidence": "5;5;4",
"confidence_avg": 4.666666666666667,
"soundness": "2;3;3",
"soundness_avg": 2.6666666666666665,
"contribution": "2;2;3",
"contribution_avg": 2.3333333333333335,
"presentation": "2;3;2",
"presentation_avg": 2.33333333333333... | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.633435"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
JJepij22fb | 2409.15360v3 | Reward-Robust RLHF in LLMs | {
"content": "## Abstract\n\nAbstract As Large Language Models continue to progress toward more advanced forms of intelligence, Reinforcement Learning from Human Feedback is increasingly seen as a key pathway toward achieving Artificial General Intelligence. However, the reliance on reward-model-based alignment metho... | [
{
"id": "AKWUpcmXAa",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "This paper presents an approach on building robust reward models (RMs) for stable RLHF. The main idea is to build a Bayesian reward model ensemble to model the un... | {
"rating": "1;3;5;6",
"rating_avg": 3.75,
"confidence": "4;4;3;3",
"confidence_avg": 3.5,
"soundness": "1;2;3;4",
"soundness_avg": 2.5,
"contribution": "1;3;2;3",
"contribution_avg": 2.25,
"presentation": "1;2;3;3",
"presentation_avg": 2.25
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.634177"
} | {
"id": "E3QUAS2Xdk",
"metareview": "The main idea is to build a Bayesian reward model ensemble to model the uncertainty set of reward functions and then run RLHF with interpolation with the standard RLHF objective and the robust RLHF objective using the minimum reward score from the RM ensemble in the uncertainty ... | {
"decision": "Reject"
} |
JLDAWbzTUg | 2405.16587v2 | Cost-Effective Online Multi-LLM Selection with Versatile Reward Models | {
"content": "## Abstract\n\nAbstract With the rapid advancement of large language models (LLMs), the diversity of multi-LLM tasks and the variability in their pricing structures have become increasingly important, as costs can vary greatly between different LLMs. To tackle these challenges, we introduce the C2MAB-V ... | [
{
"id": "tOrxpR2yc8",
"initial_rating": 6,
"confidence": 2,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "Main problem statement: the problem statement is clear and well-motivated (i.e. the setup of selecting a subset of LLMs/experts to answer a given query, in a cost... | {
"rating": "3;5;6;6",
"rating_avg": 5,
"confidence": "3;4;4;2",
"confidence_avg": 3.25,
"soundness": "2;2;3;3",
"soundness_avg": 2.5,
"contribution": "2;2;3;3",
"contribution_avg": 2.5,
"presentation": "2;2;3;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.635233"
} | {
"id": "wrVXBdXbYo",
"metareview": "The paper proposes a new algorithm for selecting a subset of large language models (LLMs) to answer a given query, considering both cost and reward. The algorithm, called C2MAB-V, is based on combinatorial bandits and attempts to balance exploration and exploitation while solvin... | {
"decision": "Reject"
} |
JMNht3SmcG | 2403.03853v3 | ShortGPT: Layers in Large Language Models are More Redundant Than You Expect | {
"content": "## Abstract\n\nAbstract As Large Language Models (LLMs) continue to advance in performance, their size has increased significantly, with current LLMs containing billions or even trillions of parameters. In this study, we identify notable redundancy across the layers of LLMs, where some layers contribute... | [
{
"id": "Ud36hWUccA",
"initial_rating": 3,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "The paper presents a layer pruning method specifically designed for LLMs. It begins by assessing layer importance using the proposed Block Influence metric, which... | {
"rating": "3;3;5;5",
"rating_avg": 4,
"confidence": "5;4;4;4",
"confidence_avg": 4.25,
"soundness": "1;2;3;3",
"soundness_avg": 2.25,
"contribution": "1;2;2;3",
"contribution_avg": 2,
"presentation": "2;2;3;2",
"presentation_avg": 2.25
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.636159"
} | {
"id": "hWgDBkr9kJ",
"metareview": "This paper studies a layer pruning method using a layer-importance score called Block Influence (BI) based on the variation in representations between input and output layers. Empirical evaluations demonstrate the effectiveness of the method.\n\nMain weaknesses mentioned by revi... | {
"decision": "Reject"
} |
JMe5FW8c3e | 2410.01643v2 | Stable Offline Value Function Learning with Bisimulation-based Representations | {
"content": "## Abstract\n\nAbstract In reinforcement learning, offline value function learning is the procedure of using an offline dataset to estimate the expected discounted return from each state when taking actions according to a fixed target policy. The stability of this procedure, i.e., whether it converges t... | [
{
"id": "xv1VrdF5Zu",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 2,
"summary": "The paper introduces Kernel Representations for Offline Policy Evaluation (KROPE), a novel algorithm aimed at stabilizing offline value function learning in reinf... | {
"rating": "3;6;6;8",
"rating_avg": 5.75,
"confidence": "5;3;3;5",
"confidence_avg": 4,
"soundness": "2;4;3;4",
"soundness_avg": 3.25,
"contribution": "2;2;3;3",
"contribution_avg": 2.5,
"presentation": "3;3;2;4",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.636988"
} | {
"id": "Eg3oPeqgrN",
"metareview": "This paper proposes learning a stable representation for least-squares value iteration by leveraging a bisimulation-style representation. However, my primary concern lies with the definition of the krope kernel introduced in the paper, which seems fundamentally flawed. The main ... | {
"decision": "Reject"
} |
JOHhktXd4a | 2410.09972v1 | Make the Pertinent Salient: Task-Relevant Reconstruction for Visual Control with Distractions | {
"content": "## Abstract\n\nAbstract Recent advancements in Model-Based Reinforcement Learning (MBRL) have made it a powerful tool for visual control tasks.\nDespite improved data efficiency, it remains challenging to train MBRL agents with generalizable perception.\nTraining in the presence of visual distractions i... | [
{
"id": "KNVIeDIbaf",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "This paper proposes Segmentation Dreamer, a model-based reinforcement learning method that extends the Dreamer framework by introducing an auxiliary task focused ... | {
"rating": "5;5;5;5;6",
"rating_avg": 5.2,
"confidence": "4;3;3;4;4",
"confidence_avg": 3.6,
"soundness": "3;2;3;2;3",
"soundness_avg": 2.6,
"contribution": "2;2;2;2;3",
"contribution_avg": 2.2,
"presentation": "3;3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.638022"
} | {
"id": "XteA1koBgC",
"metareview": "This paper introduces Segmentation Dreamer (SD), an extension of the DREAMER framework designed to improve Model-Based Reinforcement Learning (MBRL) in visually distracting environments. SD focuses representation learning on task-relevant components of the visual input by using ... | {
"decision": "Reject"
} |
JQrBYfD2gg | 2407.11098v3 | Inertial Confinement Fusion Forecasting via Large Language Models | {
"content": "## Abstract\n\nAbstract Controlled fusion energy is deemed pivotal for the advancement of human civilization. In this study, we introduce LPI-LLM , a novel integration of Large Language Models (LLMs) with classical reservoir computing paradigms tailored to address a critical challenge, Laser-Plasma Inst... | [
{
"id": "g8zkv16A6q",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "This paper presents a method to predict hard X-ray (HXR) energies emitted by the hot electrons in ICF implosions in nuclear physics. The method relies on engineer... | {
"rating": "3;5;6",
"rating_avg": 4.666666666666667,
"confidence": "4;3;3",
"confidence_avg": 3.3333333333333335,
"soundness": "2;3;3",
"soundness_avg": 2.6666666666666665,
"contribution": "2;2;3",
"contribution_avg": 2.3333333333333335,
"presentation": "3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.638915"
} | {
"id": "Zwkaw1FAOq",
"metareview": "Thank you for your submission to ICLR. This paper presents LPI-LLM, a method integrating large language models with reservoir computing, for the task of forecasting of laser plasma instabilities for inertial confinement fusion. \n\nThis is a borderline submission. Reviewers agr... | {
"decision": "Reject"
} |
JSB171dSUU | 2410.10626v1 | Efficiently Democratizing Medical LLMs for 50 Languages via a Mixture of Language Family Experts | {
"content": "## Abstract\n\nAbstract Adapting medical Large Language Models to local languages can reduce barriers to accessing healthcare services, but data scarcity remains a significant challenge, particularly for low-resource languages. To address this, we first construct a high-quality medical dataset and condu... | [
{
"id": "zzrCZpNvB7",
"initial_rating": 5,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 2,
"summary": "This paper addresses the challenge of adapting medical large language models (LLMs) to low-resource languages, where data scarcity hinders equitable access to hea... | {
"rating": "5;6;6;6",
"rating_avg": 5.75,
"confidence": "4;4;2;3",
"confidence_avg": 3.25,
"soundness": "3;4;3;3",
"soundness_avg": 3.25,
"contribution": "3;2;3;3",
"contribution_avg": 2.75,
"presentation": "2;4;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.639777"
} | {
"id": "RGLQQ4TxB3",
"metareview": "This paper tackles the task of adapting medical LLMs to new languages in order to increase access to these technologies. To do this, the work first introduces a new medical QA dataset for model tuning in 12 languages; this dataset is automatically converted from text to QA forma... | {
"decision": "Accept (Poster)"
} |
JT53iXH7eO | 2407.16982v1 | Diffree: Text-Guided Shape Free Object Inpainting with Diffusion Model | {
"content": "## Abstract\n\nAbstract This paper addresses an important problem of object addition for images with only text guidance. It is challenging because the new object must be integrated seamlessly into the image with consistent visual context, such as lighting, texture, and spatial location. While existing t... | [
{
"id": "2idj92odeX",
"initial_rating": 5,
"confidence": 5,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "This paper proposes solving the object painting problem with only text as guidance and does not rely on other constraints, e.g., shape mask. The authors first con... | {
"rating": "3;5;5;5",
"rating_avg": 4.5,
"confidence": "3;4;5;5",
"confidence_avg": 4.25,
"soundness": "2;2;3;2",
"soundness_avg": 2.25,
"contribution": "2;2;3;2",
"contribution_avg": 2.25,
"presentation": "2;2;3;2",
"presentation_avg": 2.25
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.640567"
} | {
"id": "psKyQENOeG",
"metareview": "(a) This paper introduces Diffree, a diffusion-based model trained on the proposed synthetic OABench dataset for text-guided object painting without additional constraints, demonstrating superior performance through comparisons.\n\n(b) Strengths: The paper presents a novel appro... | {
"decision": "Reject"
} |
JTji0Jfh5a | 2410.05782v1 | Reinforcement Learning from Imperfect Corrective Actions and Proxy Rewards | {
"content": "## Abstract\n\nAbstract In practice, reinforcement learning (RL) agents are often trained with a possibly imperfect proxy reward function, which may lead to a human-agent alignment issue\n(i.e., the learned policy either converges to non-optimal performance with low cumulative rewards, or achieves high ... | [
{
"id": "5roeCgfBA0",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "In this paper, a framework is proposed where a human labeler can offer additional feedback in the form of corrective actions, reflecting the labeler's action pref... | {
"rating": "3;5;6;6",
"rating_avg": 5,
"confidence": "4;3;5;4",
"confidence_avg": 4,
"soundness": "3;3;3;3",
"soundness_avg": 3,
"contribution": "2;2;3;3",
"contribution_avg": 2.5,
"presentation": "3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.641244"
} | {
"id": "GGHJgXVqK6",
"metareview": "This paper proposes that human labelers can provide corrective actions and proxy rewards. This is interesting and useful for improving RL algorithms. The algorithm's pipeline is reasonable. The experiments are effective. This algorithm has three phases: (1) human labeling of dem... | {
"decision": "Accept (Poster)"
} |
JVkdSi7Ekg | 2410.00371v1 | AHA: A Vision-Language-Model for Detecting and Reasoning Over Failures in Robotic Manipulation | {
"content": "## Abstract\n\nAbstract Robotic manipulation in open-world settings requires not only task execution but also the ability to detect and learn from failures. While recent advances in vision-language models (VLMs) and large language models (LLMs) have improved robots’ spatial reasoning and problem-solving... | [
{
"id": "BOZH6XcQLw",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This work targets the task of open-world failure detection and reasoning under the scenario of robotic manipulation, which is critical and valuable for diverse do... | {
"rating": "5;5;6;6",
"rating_avg": 5.5,
"confidence": "3;3;5;3",
"confidence_avg": 3.5,
"soundness": "2;2;3;3",
"soundness_avg": 2.5,
"contribution": "3;4;3;3",
"contribution_avg": 3.25,
"presentation": "3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.641940"
} | {
"id": "6fJVZMPKBr",
"metareview": "The paper introduces AHA, a vision-language model for detecting and reasoning about failures in robotic manipulation tasks. The method demonstrates promising performance in analyzing manipulation failures and improving downstream robotic applications, receiving generally positiv... | {
"decision": "Accept (Poster)"
} |
JWtrk7mprJ | 2411.00161v1 | Residual Deep Gaussian Processes on Manifolds | {
"content": "## Abstract\n\nAbstract We propose practical deep Gaussian process models on Riemannian manifolds, similar in spirit to residual neural networks.\nWith manifold-to-manifold hidden layers and an arbitrary last layer, they can model manifold- and scalar-valued functions, as well as vector fields.\nWe targ... | [
{
"id": "mU8mtX4fel",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 4,
"summary": "The authors propose a technique for manifold->manifold GP regression, which they propose to use as a component for deep GP regression on manifolds. They illustrat... | {
"rating": "6;6;6;8;8",
"rating_avg": 6.8,
"confidence": "4;3;3;4;2",
"confidence_avg": 3.2,
"soundness": "3;3;3;4;4",
"soundness_avg": 3.4,
"contribution": "3;2;3;4;3",
"contribution_avg": 3,
"presentation": "4;3;4;3;4",
"presentation_avg": 3.6
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Oral",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.642652"
} | {
"id": "FzdLo3TQw4",
"metareview": "The paper 'Residual Deep Gaussian Processes on Manifolds' was reviewed by 5 reviewers who gave it an average score of 7.6 (final scores: 6+8+8+8+8). The reviewers appreciated the presentation, even if they also pointed out that it is rather dense and technical. The experiments w... | {
"decision": "Accept (Oral)"
} |
JYTQ6ELUVO | 2411.02796v1 | Specialized Foundation Models struggle to beat Supervised Baselines | {
"content": "## Abstract\n\nAbstract Following its success for vision and text, the “foundation model” (FM) paradigm—pretraining large models on massive data, then fine-tuning on target tasks—has rapidly expanded to domains in the sciences, engineering, healthcare, and beyond.\nHas this achieved what the original FM... | [
{
"id": "OJX6LvImgW",
"initial_rating": 5,
"confidence": 5,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper aims to challenge the assumption that foundation models (FMs) consistently outperform traditional supervised learning methods. To support this, the aut... | {
"rating": "5;5;5;6",
"rating_avg": 5.25,
"confidence": "3;4;5;4",
"confidence_avg": 4,
"soundness": "2;3;2;4",
"soundness_avg": 2.75,
"contribution": "2;2;3;3",
"contribution_avg": 2.5,
"presentation": "3;2;3;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.643752"
} | {
"id": "orn52muh0A",
"metareview": "In this paper, the authors consider three modalities and compare \"foundational model\" performance to a standard supervised learning workflow. They show that the latter can beat the former in all modalities considered. This is an interesting piece of work that sanity checks th... | {
"decision": "Accept (Poster)"
} |
JYV2hrtFSv | 2407.06189v1 | Video-STaR: Self-Training Enables Video Instruction Tuning with Any Supervision | {
"content": "## Abstract\n\nAbstract The performance of Large Vision Language Models (LVLMs) is dependent on the size and quality of their training datasets.\nExisting video instruction tuning datasets lack diversity as they are derived by prompting large language models with video captions to generate question-answ... | [
{
"id": "A8pJl6AdFC",
"initial_rating": 6,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 2,
"summary": "The paper introduces a novel self-training approach, Video Self-Training with augmented Reasoning (Video-STaR), to improve the performance of Large Multi-modal Mo... | {
"rating": "5;5;5;6",
"rating_avg": 5.25,
"confidence": "3;3;4;4",
"confidence_avg": 3.5,
"soundness": "2;2;3;2",
"soundness_avg": 2.25,
"contribution": "3;3;2;2",
"contribution_avg": 2.5,
"presentation": "3;3;2;2",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.645311"
} | {
"id": "QyZnFM1EuO",
"metareview": "This paper proposes a self-training approach for instructing tuning MLLMs with video. \n\nThe reviewers find the approach interesting and also appreciate the dataset release. \n\nReviewers raised questions and concerns primarily with the experimental details and qualified base... | {
"decision": "Accept (Poster)"
} |
JZCxlrwjZ8 | 2405.17031v1 | Any-step Dynamics Model Improves Future Predictions for Online and Offline Reinforcement Learning | {
"content": "## Abstract\n\nAbstract Model-based methods in reinforcement learning offer a promising approach to enhance data efficiency by facilitating policy exploration within a dynamics model. However, accurately predicting sequential steps in the dynamics model remains a challenge due to the bootstrapping predi... | [
{
"id": "E3crZY0qYv",
"initial_rating": 6,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper introduces the Any-Step Dynamics Model (ADM), an RNN-based dynamics model with a variable prediction horizon. ADM enhances the fidelity of dynamics mod... | {
"rating": "3;5;5;6",
"rating_avg": 4.75,
"confidence": "3;4;3;4",
"confidence_avg": 3.5,
"soundness": "2;2;2;3",
"soundness_avg": 2.25,
"contribution": "2;2;2;3",
"contribution_avg": 2.25,
"presentation": "1;2;3;3",
"presentation_avg": 2.25
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.646180"
} | {
"id": "jOqyByiUTh",
"metareview": "This paper proposes the Any-step Dynamics Model (ADM), which is a model-based method that learns to predict transitions multiple steps into the future to avoid accumulating errors due to bootstrapping. They provide an online and an offline variant of their method, with their eff... | {
"decision": "Accept (Poster)"
} |
JaRihIHbZm | 2410.10076v2 | VideoAgent: Self-Improving Video Generation | {
"content": "## Abstract\n\nAbstract Video generation has been used to generate visual plans for controlling robotic systems. Given an image observation and a language instruction, previous work has generated video plans which are then converted to robot controls to be executed. However, a major bottleneck in levera... | [
{
"id": "XubCPvZbr1",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper introduces VideoAgent, an approach that refines video plans generated by video diffusion models through self-conditioning consistency and feedback from... | {
"rating": "3;3;5;6",
"rating_avg": 4.25,
"confidence": "4;4;4;3",
"confidence_avg": 3.75,
"soundness": "3;2;3;3",
"soundness_avg": 2.75,
"contribution": "3;2;3;3",
"contribution_avg": 2.75,
"presentation": "4;2;2;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.647029"
} | {
"id": "gsqKSPhaHe",
"metareview": "This paper proposes VideoAgent, a method for self-improving video generation. It refines video plans generated by video diffusion models through self-conditioning consistency and feedback from pre-trained vision-language models (VLMs). The refined video plans can be converted to... | {
"decision": "Reject"
} |
JeLqFpFzwX | 2406.07843v2 | Self-Attention-Based Contextual Modulation Improves Neural System Identification | {
"content": "## Abstract\n\nAbstract Convolutional neural networks (CNNs) have been shown to be state-of-the-art models for visual cortical neurons. Cortical neurons in the primary visual cortex are sensitive to contextual information mediated by extensive horizontal and feedback connections. Standard CNNs integrate... | [
{
"id": "wdz4SZC96y",
"initial_rating": 8,
"confidence": 5,
"soundness": 4,
"contribution": 3,
"presentation": 4,
"summary": "The manuscript asks if self-attention can improve predictions of primary visual cortex (V1) response to images. This is motivated by two facts: one is surround mo... | {
"rating": "3;6;8;8",
"rating_avg": 6.25,
"confidence": "4;5;4;5",
"confidence_avg": 4.5,
"soundness": "2;3;3;4",
"soundness_avg": 3,
"contribution": "2;4;3;3",
"contribution_avg": 3,
"presentation": "2;2;3;4",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.648151"
} | {
"id": "noGNph7vgV",
"metareview": "This paper describes deep learning model for predicting neural activity in primary visual cortex that incorporates convolutional layers as well as self-attention. It shows that self-attention can indeed improve performance relative to feed-forward neural networks, particularly w... | {
"decision": "Accept (Poster)"
} |
JeZS4jQF77 | 2410.17159v1 | LiNo: Advancing Recursive Residual Decomposition of Linear and Nonlinear Patterns for Robust Time Series Forecasting | {
"content": "## Abstract\n\nAbstract Forecasting models are pivotal in a data-driven world with vast volumes of time series data that appear as a compound of vast Li near and No nlinear patterns.\nRecent deep time series forecasting models struggle to utilize seasonal and trend decomposition to separate the entangle... | [
{
"id": "LYKD0KeoFD",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper presents a new neural network architecture for multi-variate time series forecasting. The main idea is to iteratively extract linear and non-linear pat... | {
"rating": "3;3;5;6",
"rating_avg": 4.25,
"confidence": "3;3;3;3",
"confidence_avg": 3,
"soundness": "3;2;2;3",
"soundness_avg": 2.5,
"contribution": "2;3;2;3",
"contribution_avg": 2.5,
"presentation": "2;1;2;3",
"presentation_avg": 2
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.649053"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
JetCx7Tpgb | 2407.08374v3 | Enhancing Robustness of Vision-Language Models through Orthogonality Learning and Self-Regularization | {
"content": "## Abstract\n\nAbstract Efficient fine-tuning of vision-language models (VLMs) like CLIP for specific downstream tasks is gaining significant attention. Previous works primarily focus on prompt learning to adapt the CLIP into a variety of downstream tasks, however, suffering from task overfitting when f... | [
{
"id": "OjZ2sJUbMg",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 2,
"summary": "This paper presents OrthSR, featuring a novel and efficient orthogonal fine-tuning method to adapt vision-language models (VLMs) to downstream tasks, along with a... | {
"rating": "3;3;5;5;5",
"rating_avg": 4.2,
"confidence": "5;5;4;4;3",
"confidence_avg": 4.2,
"soundness": "3;3;3;2;3",
"soundness_avg": 2.8,
"contribution": "2;2;2;2;2",
"contribution_avg": 2,
"presentation": "3;3;3;3;2",
"presentation_avg": 2.8
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.649839"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
JgSbMcgd8q | 2408.13233v2 | Multi-Layer Transformers Gradient Can be Approximated in Almost Linear Time | {
"content": "## Abstract\n\nThe computational complexity of the self-attention mechanism in popular transformer architectures poses significant challenges for training and inference, and becomes the bottleneck for long inputs. Is it possible to significantly reduce the quadratic time complexity of computing the grad... | [
{
"id": "oHHz2C3C21",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 2,
"summary": "The paper addresses the important problem of reducing the quadratic computational complexity of gradient calculation in transformers, proposing an approximation a... | {
"rating": "3;5;5;6",
"rating_avg": 4.75,
"confidence": "3;3;4;3",
"confidence_avg": 3.25,
"soundness": "2;3;3;3",
"soundness_avg": 2.75,
"contribution": "2;2;3;2",
"contribution_avg": 2.25,
"presentation": "2;3;3;2",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.651823"
} | {
"id": "QJzGVOxPXu",
"metareview": "The reviewers and AC found this work to be an interesting contribution. In terms of rating, this paper was borderline, slightly below the acceptance threshold. However, the final decision is rejection for the following reasons:\n\n- As initially raised by Reviewer o3CW, this wor... | {
"decision": "Reject"
} |
JiX2DuTkeU | 2410.06885v2 | F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching | {
"content": "## Abstract\n\nAbstract This paper introduces F5-TTS, a fully non-autoregressive text-to-speech system based on flow matching with Diffusion Transformer (DiT).\nWithout requiring complex designs such as duration model, text encoder, and phoneme alignment, the text input is simply padded with filler toke... | [
{
"id": "wF7ScCVAn4",
"initial_rating": 5,
"confidence": 5,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "The work is an incremental engineering effort based on E2-TTS.It replaces E2-TTS's original Transformer + U-Net structure with a Diffusion Transformer. Additional... | {
"rating": "3;5;5;8",
"rating_avg": 5.25,
"confidence": "5;4;5;3",
"confidence_avg": 4.25,
"soundness": "2;3;3;3",
"soundness_avg": 2.75,
"contribution": "1;2;2;3",
"contribution_avg": 2,
"presentation": "2;3;3;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.653693"
} | {
"id": "JwrBDj4MJy",
"metareview": "The paper proposes to use a different embedding model and a different sampling procedure to improve a nonautoregressive synthesis model, E2 TTS.\n\nI recommend a rejection because, despite the improvement and impressive performance, the root cause of the improvement is unclear. ... | {
"decision": "Reject"
} |
Jjr2Odj8DJ | 2411.06037v1 | Sufficient Context: A New Lens on Retrieval Augmented Generation Systems | {
"content": "## 1 Prompts\n\n### \\thesubsection FLAMe Prompt\n\n[colback=green!10!white, colframe=blue!75!black]\nINSTRUCTIONS:\ntitle: Is the context sufficient to infer the answer to the question?\ndescription: In this task, you will be provided with documents and a question. Use one of the following labels under... | [
{
"id": "NOXnojZLeF",
"initial_rating": 8,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This work analyzes how LLM behavior differs for question answering with and without sufficient context. A context C is considered sufficient to answer a question ... | {
"rating": "3;6;8;8",
"rating_avg": 6.25,
"confidence": "3;4;3;3",
"confidence_avg": 3.25,
"soundness": "3;3;3;3",
"soundness_avg": 3,
"contribution": "1;3;3;3",
"contribution_avg": 2.5,
"presentation": "2;3;3;4",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.658511"
} | {
"id": "aLq7ogg6S0",
"metareview": "Summary of the paper: This paper investigates the behavior of LLMs in RAG systems, focusing on the concept of \"sufficient context\"—defined as context that unambiguously supports an answer to a given question. The authors develop an \"autorater\" prompt to evaluate context suff... | {
"decision": "Accept (Poster)"
} |
JkCJBoNUcU | 2406.07255v3 | Towards Realistic Data Generation for Real-World Super-Resolution | {
"content": "## Abstract\n\nAbstract Existing image super-resolution (SR) techniques often fail to generalize effectively in complex real-world settings due to the significant divergence between training data and practical scenarios. To address this challenge, previous efforts have either manually simulated intricat... | [
{
"id": "uJUJkWbCde",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 3,
"presentation": 3,
"summary": "This paper introduces an unsupervised learning framework designed to generate realistic LR images for real-world SR. This paper utilizes a decoupled approach for ... | {
"rating": "5;5;5;8",
"rating_avg": 5.75,
"confidence": "5;4;4;5",
"confidence_avg": 4.5,
"soundness": "3;2;2;3",
"soundness_avg": 2.5,
"contribution": "3;2;3;4",
"contribution_avg": 3,
"presentation": "3;3;3;2",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.660324"
} | {
"id": "ESvmp76h4w",
"metareview": "This paper proposes a contrastive learning method to generate data for realistic image super-resolution in an unpaired setting. The image content and the degradation are decoupled to generate realistic and diverse LR images at scale. The decoupling is guided by the learning obj... | {
"decision": "Accept (Poster)"
} |
Jlhq0zb76Q | 2410.02710v1 | SteerDiff: Steering towards Safe Text-To-Image Diffusion Models | {
"content": "## Abstract\n\nAbstract Text-to-image (T2I) diffusion models have drawn attention for their ability to generate high-quality images with precise text alignment. However, these models can also be misused to produce inappropriate content. Existing safety measures, which typically rely on text classifiers ... | [
{
"id": "kkztlGtw9V",
"initial_rating": 3,
"confidence": 3,
"soundness": 1,
"contribution": 2,
"presentation": 2,
"summary": "This paper proposes a framework for removing unsafe content in the text encoder module of Stable Diffusion. This work generates a set of inappropriate concepts an... | {
"rating": "3;3;5;5",
"rating_avg": 4,
"confidence": "4;3;3;4",
"confidence_avg": 3.5,
"soundness": "2;1;2;3",
"soundness_avg": 2,
"contribution": "2;2;3;2",
"contribution_avg": 2.25,
"presentation": "2;2;3;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.661381"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
JmXu4fk5Mm | 2406.05543v1 | VP-LLM: Text-Driven 3D Volume Completion with Large Language Models through Patchification | {
"content": "## Abstract\n\nAbstract Recent conditional 3D completion works have mainly relied on CLIP or BERT to encode textual information, which cannot support complex instruction. Meanwhile, large language models (LLMs) have shown great potential in multi-modal understanding and generation tasks. Inspired by the... | [
{
"id": "9PYt6QOk9m",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "The paper presents VP-LLM, which leverages Large Language Models (LLMs) to achieve text-guided 3D volume completion. Traditional methods for 3D completion, typica... | {
"rating": "3;5;5;6",
"rating_avg": 4.75,
"confidence": "4;3;3;3",
"confidence_avg": 3.25,
"soundness": "3;3;3;3",
"soundness_avg": 3,
"contribution": "1;3;3;3",
"contribution_avg": 2.5,
"presentation": "3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.663814"
} | {
"id": "KZ7tND6E2j",
"metareview": "The paper introduces the VP-LLM, which leverages large language models (LLMs) for text-guided 3D volume completion. Several notable limitations are raised. The results were primarily based on ShapeNet with limited comparisons to baseline methods. The low-resolution outputs weake... | {
"decision": "Reject"
} |
JnWJbrnaUE | 2401.15884v3 | Corrective Retrieval Augmented Generation | {
"content": "## Abstract\n\nAbstract Large language models (LLMs) inevitably exhibit hallucinations since the accuracy of generated texts cannot be secured solely by the parametric knowledge they encapsulate.\nAlthough retrieval-augmented generation (RAG) is a practicable complement to LLMs, it relies heavily on the... | [
{
"id": "qAzStvuhJu",
"initial_rating": 3,
"confidence": 5,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "This paper proposes a CRAG framework to enhance the quality of retrieved documents in Retrieval-Augmented Generation (RAG). Specifically, a small LLM is employed ... | {
"rating": "3;3;3;6",
"rating_avg": 3.75,
"confidence": "4;5;5;4",
"confidence_avg": 4.5,
"soundness": "2;1;2;3",
"soundness_avg": 2,
"contribution": "1;2;2;2",
"contribution_avg": 1.75,
"presentation": "2;2;3;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.664757"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
Jprs1v2wPA | 2405.16124v2 | Unsupervised Meta-Learning via In-Context Learning | {
"content": "## Abstract\n\nAbstract Unsupervised meta-learning aims to learn feature representations from unsupervised datasets that can transfer to downstream tasks with limited labeled data.\nIn this paper, we propose a novel approach to unsupervised meta-learning that leverages the generalization abilities of in... | [
{
"id": "0bXlxp4XCw",
"initial_rating": 6,
"confidence": 5,
"soundness": 4,
"contribution": 3,
"presentation": 3,
"summary": "This work aims to tackle few-shot learning (FSL) by learning from unsupervised dataset. It shares a similar idea with CAML[1] in model design by taking FSL as a s... | {
"rating": "3;5;5;6",
"rating_avg": 4.75,
"confidence": "3;3;2;5",
"confidence_avg": 3.25,
"soundness": "4;3;3;2",
"soundness_avg": 3,
"contribution": "3;2;3;3",
"contribution_avg": 2.75,
"presentation": "3;3;2;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.665701"
} | {
"id": "79MPJI2Wnm",
"metareview": "This work addresses the challenge of few-shot learning (FSL) by leveraging an unsupervised dataset. It adopts a similar approach to CAML [1] in its model design, framing FSL as a sequence modeling problem. This allows the transformer encoder to learn task context from support im... | {
"decision": "Accept (Poster)"
} |
JrpMlotoGX | 2410.22257v1 | FactBench: A Dynamic Benchmark for In-the-Wild Language Model Factuality Evaluation | {
"content": "## Abstract\n\nAbstract Language models (LMs) are widely used by an increasing number of users, underscoring the challenge of maintaining factuality across a broad range of topics. We first present VERIFY ( V erification and E vidence R etr I eval for F actualit Y evaluation), a pipeline to evaluate LMs... | [
{
"id": "rE8cnMx4zg",
"initial_rating": 5,
"confidence": 4,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "This paper presents FACTBENCH, a dynamic benchmark dataset for evaluating the factuality of language model (LM) responses in real-world user interactions. The aut... | {
"rating": "5;5;5;6",
"rating_avg": 5.25,
"confidence": "4;3;4;3",
"confidence_avg": 3.5,
"soundness": "2;2;3;3",
"soundness_avg": 2.5,
"contribution": "3;3;2;3",
"contribution_avg": 2.75,
"presentation": "2;3;3;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.666824"
} | {
"id": "gDxLjWweoS",
"metareview": "This paper introduces FACTBENCH, a dynamic benchmark aimed at evaluating the factuality of language model responses by identifying and categorizing hallucinations. The research addresses an important challenge in LLM development by proposing VERIFY, a pipeline that assesses resp... | {
"decision": "Reject"
} |
Jrt9iWalFy | 2410.02388v1 | Boosting Perturbed Gradient Ascent for Last-Iterate Convergence in Games | {
"content": "## Abstract\n\nAbstract This paper presents a payoff perturbation technique, introducing a strong convexity to players’ payoff functions in games. This technique is specifically designed for first-order methods to achieve last-iterate convergence in games where the gradient of the payoff functions is mo... | [
{
"id": "v2myp5tWcl",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 2,
"summary": "This paper considers multi-player monotone games and proposes a perturbed gradient ascent algorithm with improved performance in both deterministic and stochastic... | {
"rating": "5;6;6;8",
"rating_avg": 6.25,
"confidence": "3;3;3;2",
"confidence_avg": 2.75,
"soundness": "2;3;3;4",
"soundness_avg": 3,
"contribution": "2;2;2;3",
"contribution_avg": 2.25,
"presentation": "3;3;2;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.669995"
} | {
"id": "FkAtT9FwjW",
"metareview": "The paper introduces a modifed perturbed scheme called GABP and show last iterate convergence with rate of 1/T and 1/T^1/7 for full feedback and noisy feedback models on monotone games. The reviewers agree that the paper has merits/novelty and all of them were positive or slight... | {
"decision": "Accept (Poster)"
} |
JslyktsKMY | 2407.01825v1 | Reevaluating Theoretical Analysis Methods for Optimization in Deep Learning | {
"content": "## Abstract\n\nAbstract There is a significant gap between our theoretical understanding of optimization algorithms used in deep learning and their practical performance. Theoretical development usually focuses on proving convergence guarantees under a variety of different assumptions, which are themsel... | [
{
"id": "ixTyE57XoV",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "This paper revisits two commonly used assumptions (convexity and smoothness) for analyzing the convergence of optimization algorithms. Traditionally, theoretical ... | {
"rating": "5;6;6;6",
"rating_avg": 5.75,
"confidence": "3;3;3;3",
"confidence_avg": 3,
"soundness": "3;3;3;3",
"soundness_avg": 3,
"contribution": "2;2;2;3",
"contribution_avg": 2.25,
"presentation": "3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.672438"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
Jszf4et48m | 2311.14542v2 | ToddlerDiffusion: Interactive Structured Image Generation with Cascaded Schrödinger Bridge | {
"content": "## Abstract\n\nAbstract Diffusion models break down the challenging task of generating data from high-dimensional distributions into a series of easier denoising steps. Inspired by this paradigm, we propose a novel approach that extends the diffusion framework into modality space, decomposing the comple... | [
{
"id": "e77bqu64u8",
"initial_rating": 5,
"confidence": 5,
"soundness": 3,
"contribution": 2,
"presentation": 2,
"summary": "This work proposes ToddlerDiffusion, a cascaded sequence of relatively small diffusion models, to gradually generate different aspects of images. Those different ... | {
"rating": "5;5;5;6",
"rating_avg": 5.25,
"confidence": "4;4;5;4",
"confidence_avg": 4.25,
"soundness": "3;3;3;3",
"soundness_avg": 3,
"contribution": "3;2;2;2",
"contribution_avg": 2.25,
"presentation": "3;3;2;4",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.674037"
} | {
"id": "3cCJTxYb1J",
"metareview": "The submission proposes ToddlerDiffusion, a cascaded sequence of diffusion models operating on different modalities such as contours, palettes, and textures, to finally generate an RGB image.\nInstead of using concatenation of different modalities as conditioning, they use the S... | {
"decision": "Accept (Poster)"
} |
JvH4jDDcG3 | 2403.02998v2 | Towards Calibrated Deep Clustering Network | {
"content": "## Abstract\n\nAbstract Deep clustering has exhibited remarkable performance; however, the over-confidence problem, i.e., the estimated confidence for a sample belonging to a particular cluster greatly exceeds its actual prediction accuracy, has been overlooked in prior research. To tackle this critical... | [
{
"id": "F9meavGE9X",
"initial_rating": 6,
"confidence": 4,
"soundness": 2,
"contribution": 3,
"presentation": 3,
"summary": "This paper addresses the over-confidence problem in deep clustering, where the predicted confidence for a sample significantly exceeds its actual accuracy. The au... | {
"rating": "3;6;6;6;6",
"rating_avg": 5.4,
"confidence": "3;2;4;4;4",
"confidence_avg": 3.4,
"soundness": "3;2;3;2;2",
"soundness_avg": 2.4,
"contribution": "2;2;3;3;3",
"contribution_avg": 2.6,
"presentation": "1;3;3;3;3",
"presentation_avg": 2.6
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.675870"
} | {
"id": "r3fMFC4N6z",
"metareview": "This paper presents a work that addresses the over-confidence problem in deep clustering. A novel calibration method is proposed. A new network initialization strategy is also introduced to improve training efficiency. Both theoretical analysis and empirical evaluations are prov... | {
"decision": "Accept (Poster)"
} |
Jw63fvX3QB | 2405.09798v2 | Many-Shot In-Context Learning in Multimodal Foundation Models | {
"content": "## Abstract\n\nAbstract Large language models are well-known to be effective at few-shot in-context learning (ICL). Recent advancements in multimodal foundation models have enabled unprecedentedly long context windows, presenting an opportunity to explore their capability to perform ICL with many more d... | [
{
"id": "hPFRAbZYnr",
"initial_rating": 5,
"confidence": 4,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "This paper investigates the many-shot in-context learning (ICL) capabilities of state-of-the-art multimodal foundation models. The authors benchmark multiple mult... | {
"rating": "3;5;5;5",
"rating_avg": 4.5,
"confidence": "3;3;3;4",
"confidence_avg": 3.25,
"soundness": "2;3;2;3",
"soundness_avg": 2.5,
"contribution": "2;3;2;2",
"contribution_avg": 2.25,
"presentation": "2;2;3;3",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.678461"
} | {
"id": "69a3WxSJ9D",
"metareview": "This paper obtains four negtive scores from all reviewers. Lacking of in-depth analysis is a common concern and the authors do not resolve this issue. Most of reviewers further point out that this paper is a technical report rather than a paper. Meanwhile, the link posted in the... | {
"decision": "Reject"
} |
JwoCs9O3QL | 2410.00296v1 | VLMGuard: Defending VLMs against Malicious Prompts via Unlabeled Data | {
"content": "## Abstract\n\nAbstract Vision-language models (VLMs) are essential for contextual understanding of both visual and textual information. However, their vulnerability to adversarially manipulated inputs presents significant risks, leading to compromised outputs and raising concerns about the reliability ... | [
{
"id": "iomSHrsfi1",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper addresses a challenge in VLM security in VLM security - detecting malicious prompts without requiring labeled data. The key innovation lies in analyzin... | {
"rating": "3;5;6;6",
"rating_avg": 5,
"confidence": "4;4;4;3",
"confidence_avg": 3.75,
"soundness": "2;3;3;3",
"soundness_avg": 2.75,
"contribution": "2;3;3;3",
"contribution_avg": 2.75,
"presentation": "2;3;3;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.680121"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
JwrnoB1tR0 | 2406.03474v1 | AD-H: Autonomous Driving with Hierarchical Agents | {
"content": "## Abstract\n\nAbstract Due to the impressive capabilities of multimodal large language models (MLLMs), recent works have focused on employing MLLM-based agents for autonomous driving in large-scale and dynamic environments.\nHowever, prevalent approaches often directly translate high-level instructions... | [
{
"id": "V34qy1H6Kk",
"initial_rating": 5,
"confidence": 4,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "This paper introduces a new paradigm of applying Multi Modality Large Language Model (MLLM) in autonomous driving motion planning. Compared to existing MLLM plann... | {
"rating": "1;3;5;5;6",
"rating_avg": 4,
"confidence": "5;5;3;4;5",
"confidence_avg": 4.4,
"soundness": "2;2;3;3;3",
"soundness_avg": 2.6,
"contribution": "1;1;2;2;3",
"contribution_avg": 1.8,
"presentation": "3;3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.685005"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
Jwtpbhheoy | 2405.15115v1 | Towards Better Understanding of In-Context Learning Ability from In-Context Uncertainty Quantification | {
"content": "## Abstract\n\nAbstract Predicting simple function classes has been widely used as a testbed for developing theory and understanding of the trained Transformer’s in-context learning (ICL) ability. In this paper, we revisit the training of Transformers on linear regression tasks, and different from all t... | [
{
"id": "0z8i3LWWgx",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper studies the in-context learning abilities of the trained transformer through the lens of uncertainty quantification. The authors train the transformer ... | {
"rating": "5;5;5",
"rating_avg": 5,
"confidence": "3;3;3",
"confidence_avg": 3,
"soundness": "3;2;3",
"soundness_avg": 2.6666666666666665,
"contribution": "2;2;3",
"contribution_avg": 2.3333333333333335,
"presentation": "3;2;3",
"presentation_avg": 2.6666666666666665
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.686494"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
JxhgSAnrsG | 2405.18361v1 | Is a 3D-Tokenized LLM the Key to Reliable Autonomous Driving? | {
"content": "## Abstract\n\nAbstract Rapid advancements in Autonomous Driving (AD) tasks turned a significant shift toward end-to-end fashion, particularly in the utilization of vision-language models (VLMs) that integrate robust logical reasoning and cognitive abilities to enable comprehensive end-to-end planning.\... | [
{
"id": "ir2DEQeNWb",
"initial_rating": 5,
"confidence": 5,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "The paper addresses the limitations of current vision-language models (VLMs) that rely on 2D tokenizers, which struggle with 3D perception essential for reliable ... | {
"rating": "5;5;5;5",
"rating_avg": 5,
"confidence": "4;5;5;5",
"confidence_avg": 4.75,
"soundness": "3;2;2;2",
"soundness_avg": 2.25,
"contribution": "2;2;3;2",
"contribution_avg": 2.25,
"presentation": "3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.689360"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
Jy17uvzNe5 | 2406.09771v1 | Block Coordinate Descent Methods for Optimization under J-Orthogonality Constraints with Applications | {
"content": "## Abstract\n\nAbstract The J-orthogonal matrix, also referred to as the hyperbolic orthogonal matrix, is a class of special orthogonal matrix in hyperbolic space, notable for its advantageous properties. These matrices are integral to optimization under J-orthogonal constraints, which have widespread a... | [
{
"id": "h6sXhSPBC9",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "This paper focuses on one special optimization problem with J-orthogonal constraints, a generalized version of orthogonality into signature matrix $J= X^TJX$. The... | {
"rating": "3;3;5;5",
"rating_avg": 4,
"confidence": "5;4;3;4",
"confidence_avg": 4,
"soundness": "2;2;3;2",
"soundness_avg": 2.25,
"contribution": "2;2;2;2",
"contribution_avg": 2,
"presentation": "1;2;3;3",
"presentation_avg": 2.25
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.693007"
} | {
"id": "1qQOcPkiFb",
"metareview": "This paper studies block coordinate descent method (BCD) for optimization with J-orthogonality constraint. The authors proposed a new method that minimizes a majorizing surrogate of the block restricted objective, which is obtained by sampling two rows and formulating the subpro... | {
"decision": "Reject"
} |
JyQYYjtO88 | 2212.02548v1 | Robustness of Quantum Algorithms for Nonconvex Optimization | {
"content": "## Abstract\n\nAbstract Recent results suggest that quantum computers possess the potential to speed up nonconvex optimization problems. However, a crucial factor for the implementation of quantum optimization algorithms is their robustness against experimental and statistical noises. In this paper, we ... | [
{
"id": "7XRwWaqd29",
"initial_rating": 8,
"confidence": 2,
"soundness": 4,
"contribution": 3,
"presentation": 4,
"summary": "This paper investigates the robustness of quantum algorithms for non-convex optimization, in the context of finding an $\\epsilon$-approximate second-order statio... | {
"rating": "5;5;8",
"rating_avg": 6,
"confidence": "3;3;2",
"confidence_avg": 2.6666666666666665,
"soundness": "2;4;4",
"soundness_avg": 3.3333333333333335,
"contribution": "3;2;3",
"contribution_avg": 2.6666666666666665,
"presentation": "1;4;4",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.697371"
} | {
"id": "n3Bz3g9hAb",
"metareview": "This is an interesting paper extensively studying the the complexity of finding an $\\epsilon$-approximation of a second-order stationary point of a non-convex objective function via noisy zeroth-order and gradient queries. A few quantum poly(d,epsilon) algorithms were provided,... | {
"decision": "Accept (Poster)"
} |
Jyh0DR4fFE | 2406.09116v2 | Injective flows for star-like manifolds | {
"content": "## Abstract\n\nAbstract Normalizing Flows (NFs) are powerful and efficient models for density estimation.\nWhen modeling densities on manifolds, NFs can be generalized to injective flows but the Jacobian determinant becomes computationally prohibitive.\nCurrent approaches either consider bounds on the l... | [
{
"id": "qRER2NN7Tl",
"initial_rating": 5,
"confidence": 4,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "The authors propose a normalizing flow technique for a specific class of manifolds in which the computation of the Jacobian determinant is both exact and efficien... | {
"rating": "5;5;5;5",
"rating_avg": 5,
"confidence": "4;3;4;4",
"confidence_avg": 3.75,
"soundness": "2;3;4;3",
"soundness_avg": 3,
"contribution": "3;2;2;2",
"contribution_avg": 2.25,
"presentation": "3;2;4;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.701185"
} | {
"id": "cB3AOSvakD",
"metareview": "The paper proposes injective flows for star-like manifolds of dimension $d-1$ in Euclidean space $\\mathbb{R}^d$ and show that for such manifolds the Jacobian determinant can be computed exactly and efficiently, with the same computational cost as Normalizing Flows (NFs).\nThe p... | {
"decision": "Accept (Poster)"
} |
JytL2MrlLT | 2407.03257v1 | Revisiting Nearest Neighbor for Tabular Data: A Deep Tabular Baseline Two Decades Later | {
"content": "## Abstract\n\nAbstract The growing success of deep learning in various domains has prompted investigations into its application to tabular data, where deep models have shown promising results compared to traditional tree-based methods. In this paper, we revisit Neighborhood Component Analysis (NCA), a ... | [
{
"id": "Akn2AF6dUb",
"initial_rating": 6,
"confidence": 5,
"soundness": 3,
"contribution": 3,
"presentation": 2,
"summary": "This study considers learning on tabular data, and proposes ModernNCA -- a deep version of the classic Neighborhood Components Analysis algorithm. Contrary to NCA... | {
"rating": "3;5;6;8;8",
"rating_avg": 6,
"confidence": "4;4;5;4;4",
"confidence_avg": 4.2,
"soundness": "2;2;3;3;3",
"soundness_avg": 2.6,
"contribution": "1;2;3;3;3",
"contribution_avg": 2.4,
"presentation": "2;3;2;3;3",
"presentation_avg": 2.6
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.704684"
} | {
"id": "WnBWEa0VLV",
"metareview": "This paper explores more traditional approaches for supervised tabular data modelling integrated with modern deep learning techniques. It uses a Neighbourhood Components Analysis approach in the latent space of a neural network which can be optimized with SGD.\n\nReviewers broug... | {
"decision": "Accept (Poster)"
} |
K1G8UKcEBO | 2405.15129v2 | ADMM for Nonsmooth Composite Optimization under Orthogonality Constraints | {
"content": "## Abstract\n\nAbstract We consider a class of structured, nonconvex, nonsmooth optimization problems under orthogonality constraints, where the objectives combine a smooth function, a nonsmooth concave function, and a nonsmooth weakly convex function. This class of problems finds diverse applications i... | [
{
"id": "sBpeVTDRMD",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 3,
"presentation": 2,
"summary": "This paper proposes an ADMM-type algorithm to solve a class of nonconvex nonsmooth composite problems under orthogonality constraints. The iteration complexity of... | {
"rating": "3;3;5;5",
"rating_avg": 4,
"confidence": "4;4;4;4",
"confidence_avg": 4,
"soundness": "3;2;3;2",
"soundness_avg": 2.5,
"contribution": "2;2;2;3",
"contribution_avg": 2.25,
"presentation": "2;2;2;2",
"presentation_avg": 2
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.706923"
} | {
"id": "t0XFc4gUVD",
"metareview": "This paper considers ADMM for nonsmooth composite optimization with orthogonality constraints. The authors claim that they obtained improved complexity result comparing with existing work. However, all reviewers found that the paper is not well-written. There are technical conce... | {
"decision": "Reject"
} |
K2Tqn8R9pu | 2409.08301v1 | Gaussian Differentially Private Human Faces Under a Face Radial Curve Representation | {
"content": "## Abstract\n\nAbstract In this paper we consider the problem of releasing a Gaussian Differentially Private (GDP) 3D human face. The human face is a complex structure with many features and inherently tied to one’s identity. Protecting this data, in a formally private way, is important yet challenging ... | [
{
"id": "qzvC6Ni5nu",
"initial_rating": 5,
"confidence": 2,
"soundness": 3,
"contribution": 2,
"presentation": 1,
"summary": "This paper introduces a method for creating Gaussian Differentially Private (GDP) representations of 3D human faces using a novel face radial curve representation... | {
"rating": "3;5;6",
"rating_avg": 4.666666666666667,
"confidence": "2;2;4",
"confidence_avg": 2.6666666666666665,
"soundness": "2;3;3",
"soundness_avg": 2.6666666666666665,
"contribution": "2;2;3",
"contribution_avg": 2.3333333333333335,
"presentation": "2;1;2",
"presentation_avg": 1.66666666666666... | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.709831"
} | {
"id": "0jq2cznqpc",
"metareview": "The paper introduces a Gaussian Differentially Private (GDP) method for 3D facial data using a novel radial curve-based representation to preserve privacy while maintaining structural fidelity. This paper proposes an innovative approach to integrating GDP with facial representat... | {
"decision": "Accept (Poster)"
} |
K2jOacHUlO | 2410.14675v1 | Enhancing Large Language Models' Situated Faithfulness to External Contexts | {
"content": "## Abstract\n\nAbstract Large Language Models (LLMs) are often augmented with external information as contexts, but this external information can sometimes be inaccurate or even intentionally misleading. We argue that robust LLMs should demonstrate situated faithfulness, dynamically calibrating their tr... | [
{
"id": "jfmQEh1PFq",
"initial_rating": 6,
"confidence": 3,
"soundness": 2,
"contribution": 3,
"presentation": 3,
"summary": "The authors propose multiple approaches for making large language models selectively reliant on either the context, or their parametric knowledge -- a framing ter... | {
"rating": "5;6;8;8",
"rating_avg": 6.75,
"confidence": "3;3;3;3",
"confidence_avg": 3,
"soundness": "3;2;4;4",
"soundness_avg": 3.25,
"contribution": "3;3;3;3",
"contribution_avg": 3,
"presentation": "3;3;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Spotlight",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.711845"
} | {
"id": "C7IkpJ8huk",
"metareview": "This paper introduces the concept of “situated faithfulness,” where large language models (LLMs) dynamically determine whether to rely on ex ternal contexts or their internal knowledge based on confidence assessments. The authors propose two approaches—Self-Guided Confidence Rea... | {
"decision": "Accept (Spotlight)"
} |
K3n5jPkrU6 | 2406.07155v1 | Scaling Large Language Model-based Multi-Agent Collaboration | {
"content": "## Abstract\n\nAbstract Pioneering advancements in large language model-powered agents have underscored the design pattern of multi-agent collaboration, demonstrating that collective intelligence can surpass the capabilities of each individual.\nInspired by the neural scaling law, which posits that incr... | [
{
"id": "nJ9QYgS476",
"initial_rating": 5,
"confidence": 4,
"soundness": 4,
"contribution": 3,
"presentation": 4,
"summary": "The paper introduces a novel multi-agent collaboration framework, MACNET, which utilizes directed acyclic graphs (DAGs) to organize agents and orchestrate their r... | {
"rating": "5;5;6;8",
"rating_avg": 6,
"confidence": "3;4;4;3",
"confidence_avg": 3.5,
"soundness": "1;4;3;4",
"soundness_avg": 3,
"contribution": "2;3;3;4",
"contribution_avg": 3,
"presentation": "3;4;4;4",
"presentation_avg": 3.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.713260"
} | {
"id": "jLX2u32vox",
"metareview": "This paper, using the MACNET framework, investigates the impact of scaling the number of collaborative agents on overall performance, revealing a collaborative scaling law. However, the experimental setup is significantly confusing, particularly in the implementation details, in... | {
"decision": "Accept (Poster)"
} |
K5yeB4dTtS | 2410.03450v1 | MLLM as Retriever: Interactively Learning Multimodal Retrieval for Embodied Agents | {
"content": "## Abstract\n\nAbstract MLLM agents demonstrate potential for complex embodied tasks by retrieving multimodal task-relevant trajectory data.\nHowever, current retrieval methods primarily focus on surface-level similarities of textual or visual cues in trajectories, neglecting their effectiveness for the... | [
{
"id": "PTyJrZdTlV",
"initial_rating": 6,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "The paper introduces a novel method, MLLM As ReTriever (MART), which enhances the performance of embodied agents by utilizing interaction data to fine-tune a mult... | {
"rating": "5;6;6;6",
"rating_avg": 5.75,
"confidence": "5;3;3;3",
"confidence_avg": 3.5,
"soundness": "2;3;2;3",
"soundness_avg": 2.5,
"contribution": "2;2;2;3",
"contribution_avg": 2.25,
"presentation": "3;3;2;3",
"presentation_avg": 2.75
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.714650"
} | {
"id": "PprSiCiTnh",
"metareview": "**Summary**\n\nThe paper proposes to use multi-modal large language models (MLLM) to retrieve trajectories for embodied tasks by fine-tuning the MLLM to judge the effectiveness of the retrieved trajectory. To capture important information in the trajectory, the paper introduces... | {
"decision": "Accept (Poster)"
} |
K9zedJlybd | 2405.14985v2 | Implicit degree bias in the link prediction task | {
"content": "## Abstract\n\nAbstract Link prediction—a task of distinguishing actual hidden edges from random unconnected node pairs—is one of the quintessential tasks in graph machine learning.\nDespite being widely accepted as a universal benchmark and a downstream task for representation learning, the validity of... | [
{
"id": "0mqqEqQpgC",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 3,
"presentation": 3,
"summary": "This paper is focused on the link prediction task and it shows how the sampling procedure applied in the evaluation of link prediction methods is biased towards h... | {
"rating": "3;5;5;6;6;8",
"rating_avg": 5.5,
"confidence": "4;5;4;4;4;3",
"confidence_avg": 4,
"soundness": "2;3;2;3;3;3",
"soundness_avg": 2.6666666666666665,
"contribution": "2;2;3;3;3;3",
"contribution_avg": 2.6666666666666665,
"presentation": "3;3;3;3;4;3",
"presentation_avg": 3.166666666666666... | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.716942"
} | {
"id": "YcddtCdQw9",
"metareview": "This paper discusses an important problem of how link prediction evaluation is defined in the context of graph representation learning research. The authors note that common methods for evaluating link prediction settings are flawed because the asymmetry between degree-biased p... | {
"decision": "Reject"
} |
KA2Rit4ky1 | 2402.16913v1 | PDETime: Rethinking Long-term Multivariate Time Series Forecasting from the Perspective of Partial Differential Equations | {
"content": "## Abstract\n\nAbstract Recent advancements in deep learning have led to the development of various models for long-term multivariate time-series forecasting (LMTF), many of which have shown promising results. Generally, the focus has been on historical-value-based models, which rely on past observation... | [
{
"id": "G3nhwnZVE8",
"initial_rating": 3,
"confidence": 3,
"soundness": 3,
"contribution": 2,
"presentation": 2,
"summary": "The paper presents PDETime, a novel approach for long-term multivariate time-series forecasting that models the series as a continuous dynamical system governed b... | {
"rating": "3;3;6;6;6",
"rating_avg": 4.8,
"confidence": "3;3;5;3;4",
"confidence_avg": 3.6,
"soundness": "2;3;3;2;3",
"soundness_avg": 2.6,
"contribution": "2;2;3;2;4",
"contribution_avg": 2.6,
"presentation": "1;2;3;3;3",
"presentation_avg": 2.4
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.718435"
} | {
"id": "iVdIg7wnoz",
"metareview": "This paper introduces PDETime, a novel approach to long-term multivariate time-series forecasting using partial differential equations (PDEs) to model the temporal dynamics of latent representations. While the perspective of treating time series as samples from a PDE is intrigui... | {
"decision": "Reject"
} |
KAWlH5pfQu | 2410.17442v1 | Detecting Adversarial Examples | {
"content": "## Abstract\n\nAbstract Deep Neural Networks (DNNs) have been shown to be vulnerable to adversarial examples. While numerous successful adversarial attacks have been proposed, defenses against these attacks remain relatively understudied. Existing defense approaches either focus on negating the effects ... | [
{
"id": "y9sDWzKtVL",
"initial_rating": 3,
"confidence": 4,
"soundness": 2,
"contribution": 1,
"presentation": 2,
"summary": "This paper proposes Layer Regression (LR), a universal and lightweight method for detecting adversarial examples in DNNs. The key innovation lies in analyzing the... | {
"rating": "1;3;3;3;5",
"rating_avg": 3,
"confidence": "5;5;4;4;5",
"confidence_avg": 4.6,
"soundness": "1;1;1;2;2",
"soundness_avg": 1.4,
"contribution": "1;1;3;1;2",
"contribution_avg": 1.6,
"presentation": "2;3;3;2;2",
"presentation_avg": 2.4
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Conference Withdrawn Submission",
"venueid": "ICLR.cc/2025/Conference/Withdrawn_Submission",
"processed_at": "2026-01-14T22:16:01.721214"
} | {
"id": "",
"metareview": "",
"additional_comments": ""
} | {
"decision": ""
} |
KBixkDNE8p | 2410.01677v3 | MIND SCRAMBLE: UNVEILING LARGE LANGUAGE MODEL PSYCHOLOGY VIA TYPOGLYCEMIA | {
"content": "## Abstract\n\nAbstract Although still in its infancy, research into the external behaviors and internal mechanisms of large language models (LLMs) has shown significant promise in addressing complex tasks in the physical world. These studies suggest that powerful LLMs, such as GPT-4, are beginning to e... | [
{
"id": "oAl32xTp7h",
"initial_rating": 3,
"confidence": 4,
"soundness": 3,
"contribution": 2,
"presentation": 2,
"summary": "This paper proposes to evaluate language models from the Typoglycemia perspective to verify how much alignment there is with humans in the emerging field of LLM P... | {
"rating": "3;3;3;3;3",
"rating_avg": 3,
"confidence": "5;5;3;4;4",
"confidence_avg": 4.2,
"soundness": "1;2;2;2;3",
"soundness_avg": 2,
"contribution": "2;2;2;2;2",
"contribution_avg": 2,
"presentation": "2;2;2;3;2",
"presentation_avg": 2.2
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.722389"
} | {
"id": "nl6bZTjFVc",
"metareview": "The contributions of this paper were not judged as novel and well grounded in literature by initial reviews -- authors did not engage in rebuttal or discussion.",
"additional_comments": "none"
} | {
"decision": "Reject"
} |
KCYDpqSpqg | 2405.15321v1 | SG-Adapter: Enhancing Text-to-Image Generation with Scene Graph Guidance | {
"content": "## Abstract\n\nAbstract Recent advancements in text-to-image generation have been propelled by the development of diffusion models and multi-modality learning. However, since text is typically represented sequentially in these models, it often falls short in providing accurate contextualization and stru... | [
{
"id": "XqJic8H3RP",
"initial_rating": 6,
"confidence": 4,
"soundness": 3,
"contribution": 3,
"presentation": 1,
"summary": "This paper proposes SGAdapter which solves the problem of using sequential representations for input text in recent text-to-image generation methods. They augment... | {
"rating": "5;5;6;6",
"rating_avg": 5.5,
"confidence": "4;4;4;4",
"confidence_avg": 4,
"soundness": "3;3;3;3",
"soundness_avg": 3,
"contribution": "2;2;3;3",
"contribution_avg": 2.5,
"presentation": "2;3;4;1",
"presentation_avg": 2.5
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.724254"
} | {
"id": "0XKT8m2NvG",
"metareview": "While both R1 and R2 viewed the paper marginally above the acceptance threshold (based on the simpleness and effectiveness of the method in correcting the incorrect contextualization in text embeddings and enhancing the structural semantics generation capabilities), they shared ... | {
"decision": "Reject"
} |
KCr2eoSiKF | 2410.01580v1 | Learning-Augmented Robust Algorithmic Recourse | {
"content": "## Abstract\n\nAbstract The widespread use of machine learning models in high-stakes domains can have a major negative impact, especially on individuals who receive undesirable outcomes. Algorithmic recourse provides such individuals with suggestions of minimum-cost improvements they can make to achieve... | [
{
"id": "X6OmT10CEd",
"initial_rating": 5,
"confidence": 4,
"soundness": 3,
"contribution": 1,
"presentation": 2,
"summary": "The authors study in this paper the notion of robust recourse. A recourse is defined in the first paragraph: it ``provides each individual who was given an undesi... | {
"rating": "3;5;5",
"rating_avg": 4.333333333333333,
"confidence": "4;4;4",
"confidence_avg": 4,
"soundness": "2;2;3",
"soundness_avg": 2.3333333333333335,
"contribution": "2;2;1",
"contribution_avg": 1.6666666666666667,
"presentation": "3;1;2",
"presentation_avg": 2
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.727251"
} | {
"id": "NRfb1wy4Au",
"metareview": "This paper aims to address the robust recourse problem, where a model might update over time invalidating the recourse guarantees. While reviewers acknowledged the importance of the problem, they raised a number of concerns that I believe the authors can try to address in the ne... | {
"decision": "Reject"
} |
KDGP8yAz5b | 2407.15176v2 | ReAttention: Training-Free Infinite Context with Finite Attention Scope | {
"content": "## Abstract\n\nAbstract The long-context capability of the Large Language Models (LLM) has made significant breakthroughs, but the maximum supported context length remains a critical bottleneck limiting their practical applications. The constraint of context length in LLMs arises from the self-attention... | [
{
"id": "baZXAAKhRE",
"initial_rating": 5,
"confidence": 5,
"soundness": 3,
"contribution": 1,
"presentation": 2,
"summary": "This paper proposes ReAttention, a a training-free approach enabling LLM based on the self-attention mechanism to support an infinite context with a finite attent... | {
"rating": "3;5;5;8",
"rating_avg": 5.25,
"confidence": "5;3;5;3",
"confidence_avg": 4,
"soundness": "2;2;3;3",
"soundness_avg": 2.5,
"contribution": "2;2;1;3",
"contribution_avg": 2,
"presentation": "2;2;2;3",
"presentation_avg": 2.25
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.729194"
} | {
"id": "JgQsPTbPoK",
"metareview": "The paper proposes ReAttention, a novel training-free method that enables large language models (LLMs) to handle effectively infinite context lengths with a finite attention scope. The key innovation lies in performing a position-agnostic top-k attention before the regular self-... | {
"decision": "Accept (Poster)"
} |
KFLtFSOtdj | 2409.19283v2 | Analyzing and Mitigating Inconsistency in Discrete Audio Tokens for Neural Codec Language Models | {
"content": "## 1 abstract\n\nBuilding upon advancements in Large Language Models (LLMs), the field of audio processing has seen increased interest in training audio generation tasks with discrete audio token sequences. However, directly discretizing audio by\nneural audio codecs often results in sequences that fund... | [
{
"id": "XXHYgHZB5o",
"initial_rating": 5,
"confidence": 5,
"soundness": 3,
"contribution": 2,
"presentation": 3,
"summary": "The paper investigates the phenomenon of Discrete Representation Inconsistency (DRI) in audio token sequences generated by neural audio codecs. Unlike text tokens... | {
"rating": "1;5;5;6;6",
"rating_avg": 4.6,
"confidence": "5;3;3;3;5",
"confidence_avg": 3.8,
"soundness": "1;3;2;3;3",
"soundness_avg": 2.4,
"contribution": "1;3;2;3;3",
"contribution_avg": 2.4,
"presentation": "2;2;2;3;3",
"presentation_avg": 2.4
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.730673"
} | {
"id": "3vuBelr9w1",
"metareview": "> The paper investigates the phenomenon of Discrete Representation Inconsistency (DRI) in audio token sequences generated by neural audio codecs. Unlike text tokens, which are deterministic, discrete audio tokens can vary significantly even if the perceptual audio remains identi... | {
"decision": "Reject"
} |
KGRV73Zcqt | 2410.04088v1 | Cross Resolution Encoding-Decoding For Detection Transformers | {
"content": "## Abstract\n\nAbstract Detection Transformers (DETR) are renowned object detection pipelines, however computationally efficient multiscale detection using DETR is still challenging. In this paper, we propose a Cross-Resolution Encoding-Decoding (CRED) mechanism that allows DETR to achieve the accuracy ... | [
{
"id": "jUsr1mPA52",
"initial_rating": 5,
"confidence": 3,
"soundness": 3,
"contribution": 3,
"presentation": 3,
"summary": "This paper introduces a Cross-Resolution Encoding-Decoding (CRED) mechanism to enhance the efficiency of DETR for object detection. CRED consists of two key modul... | {
"rating": "5;5;6;6",
"rating_avg": 5.5,
"confidence": "4;3;3;3",
"confidence_avg": 3.25,
"soundness": "3;3;3;4",
"soundness_avg": 3.25,
"contribution": "2;3;2;3",
"contribution_avg": 2.5,
"presentation": "3;3;2;4",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.731321"
} | {
"id": "oaIByn9SQz",
"metareview": "To improve the efficiency of DETR computation, the Cross-Resolution Encoding-Decoding (CRED) mechanism has been proposed. CRED uses two modules, CRAM and OSMA, to achieve both low-resolution speed and high-resolution accuracy. CRAM transfers low-resolution information to high-re... | {
"decision": "Reject"
} |
KGZAs8VcOM | 2406.10163v2 | MeshAnything: Artist-Created Mesh Generation with Autoregressive Transformers | {
"content": "## Abstract\n\nAbstract Recently, 3D assets created via reconstruction and generation have matched the quality of manually crafted assets, highlighting their potential for replacement.\nHowever, this potential is largely unrealized because these assets always need to be converted to meshes for 3D indust... | [
{
"id": "QiWagAuoZA",
"initial_rating": 5,
"confidence": 3,
"soundness": 2,
"contribution": 2,
"presentation": 3,
"summary": "This work introduces a method for conditional \"artist-created\" mesh generation. \nThe key idea is employing auto-regressive transformer on top of a VQ-VAE pre-t... | {
"rating": "5;5;6;8",
"rating_avg": 6,
"confidence": "5;3;5;4",
"confidence_avg": 4.25,
"soundness": "3;2;3;3",
"soundness_avg": 2.75,
"contribution": "3;2;3;4",
"contribution_avg": 3,
"presentation": "4;3;3;4",
"presentation_avg": 3.5
} | {
"primary_area": "",
"track": "main",
"venue": "ICLR 2025 Poster",
"venueid": "ICLR.cc/2025/Conference",
"processed_at": "2026-01-14T22:16:01.732885"
} | {
"id": "0Dfg3QoiWG",
"metareview": "The submission received somewhat mixed reviews. The reviewers appreciate the work addressing artist-like shape generation for practical artistic applications, and strong results were presented. The major concerns from 7tHF and tUZi were on doubts on scalability and missing/unfai... | {
"decision": "Accept (Poster)"
} |
KHTkRhq2aB | 2406.17923v1 | PAFT: A Parallel Training Paradigm for Effective LLM Fine-Tuning | {
"content": "## Abstract\n\nAbstract Large language models (LLMs) have shown remarkable abilities in diverse natural language processing (NLP) tasks. The LLMs generally undergo supervised fine-tuning (SFT) followed by preference alignment to be usable in downstream applications. However, this sequential training pip... | [
{
"id": "0jDZzZKhOK",
"initial_rating": 5,
"confidence": 4,
"soundness": 2,
"contribution": 3,
"presentation": 4,
"summary": "This paper proposes a novel training approach to perform parallel training of SFT and preference alignment (like PPO, DPO). Their approach aims at solving the pro... | {
"rating": "3;5;6;8",
"rating_avg": 5.5,
"confidence": "4;4;4;3",
"confidence_avg": 3.75,
"soundness": "3;2;3;3",
"soundness_avg": 2.75,
"contribution": "2;3;3;3",
"contribution_avg": 2.75,
"presentation": "2;4;3;3",
"presentation_avg": 3
} | {
"primary_area": "",
"track": "main",
"venue": "Submitted to ICLR 2025",
"venueid": "ICLR.cc/2025/Conference/Rejected_Submission",
"processed_at": "2026-01-14T22:16:01.734135"
} | {
"id": "MvyhMh0Z4W",
"metareview": "The paper proposes a novel training method for fine-tuning large language models (LLMs). The approach independently trains a model using supervised fine-tuning (SFT) and preference alignment methods like DPO and then merges them using sparsity-aware techniques to address redunda... | {
"decision": "Reject"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.