year int64 2.03k 2.03k | id stringlengths 10 10 | rating listlengths 0 9 | decision stringclasses 1 value | reviewer_comments listlengths 0 9 | _raw_metadata dict |
|---|---|---|---|---|---|
2,026 | 00F7BfXLYJ | [
4,
4,
4,
4
] | [
{
"content": "This paper addresses the limitations of current Multimodal Large Language Models (MLLMs) in deep logical reasoning for video understanding—such as feed-forward processing constraints (lack of self-correction), poor test-time scaling, and hallucinations. Inspired by cybernetic principles (control, ... | {
"cdate": 1757998013559,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025cyberv,\ntitle={CyberV: A Cybernetic Framework for Enhancing Logical Reasoning in Video Understanding},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=00F7BfXLYJ},\nnote={under review}\n}"
},
"abstract": {
"value": "Current Multimodal Large Language Models (MLLMs) may struggle with tasks requiring deep logical reasoning about video content, primarily stemming from the feed-forward processing nature, which limits their ability for self-correction and iterative refinement. To address these limitations, we propose a novel framework inspired by cybernetic principles, redesigning video MLLMs as adaptive systems capable of self-monitoring, self-correction, and dynamic resource allocation during inference. Our approach, CyberV, introduces a cybernetic loop consisting of an MLLM Inference System, a Sensor, and a Controller. Specifically, the sensor monitors MLLM forward processes. It collects intermediate interpretations, such as attention drift, then the controller determines when and how to trigger self-correction and generate feedback to guide the next round. This test-time adaptive scaling framework enhances frozen MLLMs without requiring training or additional components. Experiments demonstrate significant improvements on complex reasoning benchmarks: CyberV boosts Qwen2.5-VL-7B by 8.3% and InternVL3-8B by 5.5% on VideoMMMU, surpassing the competitive proprietary model GPT-4o. When applied to Qwen2.5-VL-72B, it yields a 10.0% improvement, achieving performance even comparable to human experts. Furthermore, on other reasoning-focused benchmarks, our method shows consistent gains of 4.6% on the multiple-choice question section of MMVU and 2.4% on MMR-V, highlighting its robustness in enhancing logical reasoning for video understanding. The code will be released to support further research."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Video Understanding",
"Multimodal Large Language Models",
"Test-Time Scaling"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/6befca6b66a747daaa91eea1475167c914c23565.pdf"
},
"primary_area": {
"value": "applications to computer vision, audio, language, and other modalities"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "CyberV: A Cybernetic Framework for Enhancing Logical Reasoning in Video Understanding"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "00F7BfXLYJ",
"id": "00F7BfXLYJ",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission6845/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897888857,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission6845/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission6845/Authors"
]
} | |
2,026 | 00HNN8O7Ni | [
4,
2,
2,
4
] | [
{
"content": "This paper proposed a new reinforcement learning framework of synthesizing hardware circuits based on the feedback from model checking results.\nThe experiments are based on open datasets and the results are outperform supervised learning baselines.\n\nPros:\n1. The integration of model checking r... | {
"cdate": 1758322705432,
"content": {
"TLDR": {
"value": "We propose a deep learning approach for reactive synthesis that first initializes a model with imitation learning and then continues training by reinforcing formally verified solutions."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025learning,\ntitle={Learning Reactive Synthesis from Model Checking Feedback},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=00HNN8O7Ni},\nnote={under review}\n}"
},
"abstract": {
"value": "Deep learning applications to formal verification typically fall into one of two categories: employing reinforcement learning that suffers from slow convergence, or supervised learning that suffers from limited exploration. For reactive synthesis, the problem of automatically constructing a system that satisfies a formal specification, existing approaches fall into the latter category. In this paper, we propose a hybrid approach that only initializes the model with supervised learning and then continues training by reinforcing formally verified predictions. We show that by training the model to synthesize correct solutions rather than fixating on the supervised data, performance substantially improves. We can further utilize our approach to optimize for size without any performance degradation. Finally, we show that we can iteratively reinforce on open problems that synthesis tools are unable to solve. Our approach is demonstrated for both deep neural networks trained from scratch and pre-trained models fine-tuned on reactive synthesis, establishing new state-of-the-art results for learning reactive synthesis."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Temporal Logic",
"Reactive Synthesis",
"Expert Iteration"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/34d3a3eeb460a6177f52996e217332dfd2836e22.pdf"
},
"primary_area": {
"value": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Learning Reactive Synthesis from Model Checking Feedback"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "00HNN8O7Ni",
"id": "00HNN8O7Ni",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission21857/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759896899730,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission21857/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission21857/Authors"
]
} | |
2,026 | 00UQtHqB2k | [
2,
6,
2,
4
] | [
{
"content": "The paper proposes a unified way to evaluate group fairness through sparsity. It studies links among Maximum Pairwise Difference, the Gini Index, and a PQ Index and argues that higher sparsity means lower fairness. Based on this view, it replaces the pairwise step in common criteria with a sparsit... | {
"cdate": 1758232139112,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025toward,\ntitle={Toward Unifying Group Fairness Evaluation from a Sparsity Perspective},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=00UQtHqB2k},\nnote={under review}\n}"
},
"abstract": {
"value": "Ensuring algorithmic fairness remains a significant challenge in machine learning, particularly as models are increasingly applied across diverse domains. While numerous fairness criteria exist, they often lack generalizability across different machine learning problems. This paper examines the connections and differences among various sparsity measures in promoting fairness and proposes a unified sparsity-based framework for evaluating algorithmic fairness. The framework aligns with existing fairness criteria and demonstrates broad applicability to a wide range of machine learning tasks. We demonstrate the effectiveness of the proposed framework as an evaluation metric through extensive experiments on a variety of datasets and bias mitigation methods. This work provides a novel perspective to algorithmic fairness by framing it through the lens of sparsity and social equity, offering potential for broader impact on fairness research and applications."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Fairness",
"Sparsity",
"Unified Framework"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/219ccddd225cef5a883ca674d9f1b6bc2e08423c.pdf"
},
"primary_area": {
"value": "alignment, fairness, safety, privacy, and societal considerations"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/fde30f02a6849cd5c614e87efe679a0e788d23bb.zip"
},
"title": {
"value": "Toward Unifying Group Fairness Evaluation from a Sparsity Perspective"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "00UQtHqB2k",
"id": "00UQtHqB2k",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission14292/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897378369,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission14292/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission14292/Authors"
]
} | |
2,026 | 017F77AYeQ | [
2,
2,
4,
0
] | [
{
"content": "The paper proposes SMART-3D, a mask token modeling approach for 3D generation.",
"id": "gZowcvNNqh",
"rating": 2
},
{
"content": "The paper proposes an framework that merges masked autoregressive generation with diffusion modeling and linear attention, addressing key efficiency bot... | {
"cdate": 1758113495159,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025smartd,\ntitle={{SMART}-3D: Scaling Masked AutoRegressive Transformer for Efficient 3D Shape Generation},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=017F77AYeQ},\nnote={under review}\n}"
},
"abstract": {
"value": "Autoregressive models have shown promise in 3D shape generation by modeling complex spatial dependencies between discrete shape tokens. However, their sequential nature and token-by-token sampling limit scalability and generation speed, especially for high-resolution shapes. In this work, we propose SMART-3D (Scaling Masked AutoRegressive Transformers for 3D generation), a novel framework that combines the modeling capacity of autoregressive transformers with the efficiency of masked generation. By introducing a hierarchical token representation and a progressive masked generation schedule, SMART-3D enables parallel decoding of 3D structures without sacrificing autoregressive fidelity. We further optimize the model with spatially-aware masking and lightweight transformer blocks, allowing generation of detailed 3D shapes with significantly reduced computational overhead. Experiments on ShapeNet, ModelNet, and ShapeNet-55 datasets demonstrate that SMART-3D achieves state-of-the-art performance in both generation quality and speed, outperforming previous competitive baselines. Our approach offers a scalable and practical solution for high-fidelity 3D shape synthesis in real-world applications."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Autoregressive models",
"3D shape generation"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/676ed3977332fe4f530434b6e3796debb83cbe57.pdf"
},
"primary_area": {
"value": "generative models"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "SMART-3D: Scaling Masked AutoRegressive Transformer for Efficient 3D Shape Generation"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "017F77AYeQ",
"id": "017F77AYeQ",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission9157/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897740443,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission9157/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission9157/Authors"
]
} | |
2,026 | 023yMrtHQP | [
4,
4,
4
] | [
{
"content": "This paper introduces a prompting framework, named Expectation–Evidence Prompting (EEP), for large language models to enhance factual verification. Drawing from the Strategic Use of Evidence technique in cognitive psychology, EEP involves generating two sets of expectations, supportive and refutat... | {
"cdate": 1758292986416,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025expectationevidence,\ntitle={Expectation{\\textendash}Evidence Prompting: Structuring Verification by Comparing Expected and Observed Evidence},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=023yMrtHQP},\nnote={under review}\n}"
},
"abstract": {
"value": "Large language models (LLMs) often fail in factual verification due to hallucinations, unreliable truthfulness judgments, and opaque reasoning. We identify a structural limitation underlying these failures: LLMs directly compare claims with evidence without accounting for expected refutational alternatives. Specifically, we demonstrate that this omission leads to ambiguity in contradiction detection and unreliable abstention. Leveraging this observation, we introduce Expectation-Evidence Prompting (EEP), a cognitively inspired strategy that first generates supportive and refutational expectations from a claim and then aligns them with observed evidence. This bidirectional reasoning process enforces logical symmetry, reduces bias toward agreement, and provides a principled abstention mechanism. Across three fact-checking benchmarks: FEVER, PubHealth, and SciFact, EEP achieves consistent gains over strong prompting baselines, including an 86.3 macro-F1 on FEVER (+3.6 over Chain-of-Thought), 82.1 precision on PubHealth (highest among all methods), and 76.1 F1 on the Supports class in SciFact. These results demonstrate that embedding expectation evidence alignment into prompt design yields more interpretable, robust, and trustworthy factual reasoning in LLMs."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Large Language Models (LLMs)",
"Factual Verification",
"Prompt Engineering",
"Cognitive Psychology–Inspired Prompting",
"Expectation–Evidence Alignment",
"Contradiction Detection",
"Abstention Mechanism"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/da7fb984ac74ee03e0b7788c1519b84d690a4cbf.pdf"
},
"primary_area": {
"value": "other topics in machine learning (i.e., none of the above)"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Expectation–Evidence Prompting: Structuring Verification by Comparing Expected and Observed Evidence"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "023yMrtHQP",
"id": "023yMrtHQP",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission19036/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897064617,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission19036/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission19036/Authors"
]
} | |
2,026 | 02NbD16OnA | [
4,
4,
4,
6
] | [
{
"content": "This paper introduces DECEPTIONDECODED, a multimodal news benchmark with explicitly defined creator intent to support misleading intent detection, source attribution, and desire inference. It reveals that current VLMs fail to reason about intent beyond surface alignment and stylistic cues.",
"... | {
"cdate": 1756910313383,
"content": {
"TLDR": {
"value": "We reveal that state-of-the-art VLMs remain blind to misleading creator intent, establishing the need for intent-aware benchmarks and models as the next frontier in multimodal misinformation detection."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025seeing,\ntitle={Seeing Through Deception: Uncovering Misleading Creator Intent in Multimodal News with Vision-Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=02NbD16OnA},\nnote={under review}\n}"
},
"abstract": {
"value": "The impact of misinformation arises not only from factual inaccuracies but also from the misleading narratives that creators deliberately embed. Interpreting such creator intent is therefore essential for multimodal misinformation detection (MMD) and effective information governance. To this end, we introduce DeceptionDecoded, a large-scale benchmark of 12,000 image–caption pairs grounded in trustworthy reference articles, created using an intent-guided simulation framework that models both the desired influence and the execution plan of news creators. The dataset captures both misleading and non-misleading cases, spanning manipulations across visual and textual modalities, and supports three intent-centric tasks: (1) misleading intent detection, (2) misleading source attribution, and (3) creator desire inference. We evaluate 14 state-of-the-art vision–language models (VLMs) and find that they struggle with intent reasoning, often relying on shallow cues such as surface-level alignment, stylistic polish, or heuristic authenticity signals. These results highlight the limitations of current VLMs and position DeceptionDecoded as a foundation for developing intent-aware models that go beyond shallow cues in MMD."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"multimodal misinformation detection",
"vision-language models",
"creator intent"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/9be01177d5da89276e95a5c85b7ef81c5e6a455e.pdf"
},
"primary_area": {
"value": "datasets and benchmarks"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Seeing Through Deception: Uncovering Misleading Creator Intent in Multimodal News with Vision-Language Models"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "02NbD16OnA",
"id": "02NbD16OnA",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission1711/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759898192988,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission1711/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission1711/Authors"
]
} | |
2,026 | 02cEkpURXH | [
2,
2,
6,
4
] | [
{
"content": "This paper proposes a KD–based training strategy for OOD generalization. The authors first argue that training compact student models via simple KD from a teacher with strong OOD performance can often surpass standalone algorithmic DG methods. They further note that prior OOD-oriented KD approache... | {
"cdate": 1758311939461,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025early,\ntitle={Early Layer Readouts for Robust Knowledge Distillation},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=02cEkpURXH},\nnote={under review}\n}"
},
"abstract": {
"value": "Domain generalization (DG) aims to learn a model that can generalize to unseen i.e. out-of-distribution (OOD) test domain. While large-capacity networks trained with sophisticated DG algorithms tend to achieve high robustness, they tend to be impractical in deployment. Typically, Knowledge distillation (KD) can alleviate this via an efficient transfer of knowledge from a robust teacher to a smaller student network. Throughout our experiments, we find that vanilla KD already provides strong OOD performance, often outperforming standalone DG algorithms. Motivated by this observation, we propose an adaptive distillation strategy that utilizes early layer predictions and uncertainty measures to learn a meta network that effectively rebalances supervised and distillation losses as per sample difficulty. Our method adds no inference overhead and consistently outperforms canonical ERM, vanilla KD, and competing DG algorithms across OOD generalization benchmarks."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"domain generalization",
"knowledge distillation",
"early layer readouts"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/2bb11bab4ab35adbf1f2a9ad3d46d601f3b0111c.pdf"
},
"primary_area": {
"value": "other topics in machine learning (i.e., none of the above)"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Early Layer Readouts for Robust Knowledge Distillation"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "02cEkpURXH",
"id": "02cEkpURXH",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission20949/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759896950334,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission20949/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission20949/Authors"
]
} | |
2,026 | 02mBAZjFzp | [
4,
4,
4,
6
] | [
{
"content": "This paper introduces VRPAGENT, a framework for discovering heuristic operators for Vehicle Routing Problems (VRPs) using large language models (LLMs). The method combines LLM-generated “destroy” and “order” operators with a Large Neighborhood Search (LNS) metaheuristic, leveraging genetic algorit... | {
"cdate": 1758296070926,
"content": {
"TLDR": {
"value": "We introduce VRPAgent, a framework that leverages LLMs and evolutionary search to discover novel heuristic operators for vehicle routing problems, achieving state-of-the-art performance across multiple VRP variants."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025vrpagent,\ntitle={{VRPA}gent: {LLM}-Driven Discovery of Heuristic Operators for Vehicle Routing Problems},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=02mBAZjFzp},\nnote={under review}\n}"
},
"abstract": {
"value": "Designing high-performing heuristics for vehicle routing problems (VRPs) is a complex task that requires both intuition and deep domain knowledge. Large language model (LLM)-based code generation has recently shown promise across many domains, but it still falls short of producing heuristics that rival those crafted by human experts. In this paper, we propose VRPAgent, a framework that integrates LLM-generated components into a metaheuristic and refines them through a novel genetic search. By using the LLM to generate problem-specific operators, embedded within a generic metaheuristic framework, VRPAgent keeps tasks manageable, guarantees correctness, and still enables the discovery of novel and powerful strategies. Across multiple problems, including the capacitated VRP, the VRP with time windows, and the prize-collecting VRP, our method discovers heuristic operators that outperform handcrafted methods and recent learning-based approaches while requiring only a single CPU core. To our knowledge, VRPAgent is the first LLM-based paradigm to advance the state-of-the-art in VRPs, highlighting a promising future for automated heuristics discovery."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"automated algorithm design",
"evolutionary search",
"vehicle routing problem",
"LLM agent",
"heuristic discovery"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/35f37aa40fad450cb00124cdc83059fbb4cb843f.pdf"
},
"primary_area": {
"value": "optimization"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "VRPAgent: LLM-Driven Discovery of Heuristic Operators for Vehicle Routing Problems"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "02mBAZjFzp",
"id": "02mBAZjFzp",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission19416/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897040045,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission19416/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission19416/Authors"
]
} | |
2,026 | 02mgFnnfqG | [
4,
8,
6,
6
] | [
{
"content": "The paper presents LiveMoments, a method for selecting and restoring a new low-quality (LQ) key photo from a short clip surrounding some key high-quality (HQ) photo. To this end, the authors build a model based on latent flow models and learnable networks for the HQ key image, the LQ candidate, an... | {
"cdate": 1757934812324,
"content": {
"TLDR": {
"value": "We are the first to restore reselected key photos in Live Photos, achieving perceptual fidelity beyond existing solutions in real-world scenes."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025livemoments,\ntitle={LiveMoments: Reselected Key Photo Restoration in Live Photos via Reference-guided Diffusion},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=02mgFnnfqG},\nnote={under review}\n}"
},
"abstract": {
"value": "Live Photo captures both a high-quality key photo and a short video clip to preserve the precious dynamics around the captured moment. \nWhile users may choose alternative frames as the key photo to capture better expressions or timing, these frames often exhibit noticeable quality degradation, as the photo capture ISP pipeline delivers significantly higher image quality than the video pipeline. This quality gap highlights the need for dedicated restoration techniques to enhance the reselected key photo. To this end, we propose LiveMoments, a reference-guided image restoration framework tailored for the reselected key photo in Live Photos. Our method employs a two-branch neural network: a reference branch that extracts structural and textural information from the original high-quality key photo, and a main branch that restores the reselected frame using the guidance provided by the reference branch. Furthermore, we introduce a unified Motion Alignment module that incorporates motion guidance for spatial alignment at both the latent and image levels. Experiments on real and synthetic Live Photos demonstrate that LiveMoments significantly improves perceptual quality and fidelity over existing solutions, especially in scenes with fast motion or complex structures."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Live Photo",
"Reference-based Image Restoration",
"Conditional Image Generation",
"Motion Alignment"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/bbbb05b5353518a72b45118dfb2eecd0c3ed7f78.pdf"
},
"primary_area": {
"value": "applications to computer vision, audio, language, and other modalities"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "LiveMoments: Reselected Key Photo Restoration in Live Photos via Reference-guided Diffusion"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "02mgFnnfqG",
"id": "02mgFnnfqG",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission5782/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897954152,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission5782/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission5782/Authors"
]
} | |
2,026 | 032sg6mGp9 | [
4,
4,
6,
6
] | [
{
"content": "This paper introduces a multinomial mixture modelling approach to address the identifiability problem in learning from noisy labels (LNL). The authors theoretically prove that LNL becomes identifiable when each sample has at least 2C−1 independent noisy labels, enabling the unique recovery of clea... | {
"cdate": 1758285923748,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025identifiability,\ntitle={Identifiability in Noisy Label Learning: A Multinomial Mixture Modelling Approach},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=032sg6mGp9},\nnote={under review}\n}"
},
"abstract": {
"value": "Learning from noisy labels (LNL) is crucial in deep learning, in which one of the approaches is to identify clean-label samples from poorly-annotated datasets. Such an identification is challenging because the conventional LNL problem, which assumes only one noisy label per instance, is non-identifiable, i.e., clean labels cannot be estimated theoretically without additional heuristics. This paper presents a novel data-driven approach that addresses this issue without requiring any heuristics about clean samples. We discover that the LNL problem becomes identifiable if there are at least $2C - 1$ i.i.d. noisy labels per instance, where $C$ is the number of classes. Our finding relies on the assumption of i.i.d. noisy labels and multinomial mixture modelling, making it easier to interpret than previous studies that require full-rank noisy-label transition matrices. To fulfil this condition without additional manual annotations, we propose a method that automatically generates additional i.i.d. noisy labels through nearest neighbours. These noisy labels are then used in the Expectation-Maximisation algorithm to infer clean labels. Our method demonstrably estimates clean labels accurately across various label noise benchmarks, including synthetic, web-controlled, and real-world datasets. Furthermore, the model trained with our method performs competitively with many state-of-the-art methods."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"label noise learning",
"expectation-maximisation",
"mixture models"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/39e718f6250a4d1ffcf2cdc9270d45e29131db80.pdf"
},
"primary_area": {
"value": "other topics in machine learning (i.e., none of the above)"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Identifiability in Noisy Label Learning: A Multinomial Mixture Modelling Approach"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "032sg6mGp9",
"id": "032sg6mGp9",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission18276/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897114753,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission18276/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission18276/Authors"
]
} | |
2,026 | 03Ek1qDZmI | [
4,
4,
4,
2
] | [
{
"content": "This paper introduces SSTP, a sample selection framework for trajectory prediction. The primary motivation is to address two challenges in existing large-scale datasets: the high computational cost of training and the imbalance where common, low-density scenarios dominate over rare, safety-critica... | {
"cdate": 1757189578927,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@misc{\nyang2025sstp,\ntitle={{SSTP}: Efficient Sample Selection for Trajectory Prediction},\nauthor={Ruining Yang and Yi Xu and Yun Fu and Lili Su},\nyear={2025},\nurl={https://openreview.net/forum?id=03Ek1qDZmI}\n}"
},
"abstract": {
"value": "Trajectory prediction is a core task in autonomous driving. However, training advanced trajectory prediction models on existing large-scale datasets is both time-consuming and computationally expensive. More critically, these datasets are highly imbalanced in scenario density, with normal driving scenes (low-moderate traffic) overwhelmingly dominating the datasets, while high-density and safety-critical cases are underrepresented. As a result, models tend to overfit low/moderate-density scenarios and perform poorly in high-density scenarios. To address these challenges, we propose the SSTP framework, which constructs a compact yet density-balanced dataset tailored to trajectory prediction. SSTP consists of two main stages: (1) Extraction, where a baseline model is pretrained for a few epochs to obtain stable gradient estimates, and the dataset is partitioned by scenario density. (2) Selection, where gradient-based scores and a submodular objective select representative samples within each density category, while biased sampling emphasizes rare high-density interactions to avoid dominance by low-density cases. This approach significantly reduces the dataset size and mitigates scenario imbalance, without sacrificing prediction accuracy. Experiments on the Argoverse 1 and Argoverse 2 datasets with recent state-of-the-art models show that SSTP achieves comparable performance to full-dataset training using only half the data while delivering substantial improvements in high-density traffic scenes and significantly reducing training time. Robust trajectory prediction depends not only on data scale but also on balancing scene density to ensure reliable performance under complex multi agent interactions. The code is available at https://anonymous.4open.science/r/SSTP_v2-69E5/README.md."
},
"anonymous_url": null,
"authorids": {
"value": [
"~Ruining_Yang1",
"~Yi_Xu9",
"~Yun_Fu1",
"~Lili_Su1"
]
},
"authors": {
"value": [
"Ruining Yang",
"Yi Xu",
"Yun Fu",
"Lili Su"
]
},
"code_of_ethics": null,
"keywords": {
"value": [
"data efficiency",
"trajectory prediction"
]
},
"no_acknowledgement_section": null,
"paperhash": {
"value": "yang|sstp_efficient_sample_selection_for_trajectory_prediction"
},
"pdf": {
"value": "/pdf/55bd982183b342ab8876bf09c69dfa0fea486112.pdf"
},
"primary_area": {
"value": "applications to robotics, autonomy, planning"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "SSTP: Efficient Sample Selection for Trajectory Prediction"
},
"venue": {
"value": "ICLR 2026 Conference Withdrawn Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Withdrawn_Submission"
}
},
"forum": "03Ek1qDZmI",
"id": "03Ek1qDZmI",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission2669/-/Full_Submission",
"ICLR.cc/2026/Conference/-/Withdrawn_Submission"
],
"license": "CC BY 4.0",
"mdate": 1762981127212,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission2669/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission2669/Authors"
]
} | |
2,026 | 03MfCNn3pF | [
2,
4,
2,
6
] | [
{
"content": "This paper presents PersonalQ, a two-stage system for personalized diffusion model serving. Check-in selects the intended personalized checkpoint via metadata reasoning and LLM-based prompt clarification, while Trigger-Aware Quantization (TAQ) preserves trigger-token features during quantization t... | {
"cdate": 1757994763056,
"content": {
"TLDR": {
"value": "PersonalQ enables efficient serving of personalized diffusion models at scale through intelligent checkpoint selection and trigger-token-aware quantization that preserves personalization quality while reducing memory footprint."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025personalq,\ntitle={PersonalQ: Select, Quantize, and Serve Personalized Diffusion Models for Efficient Inference},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=03MfCNn3pF},\nnote={under review}\n}"
},
"abstract": {
"value": "Personalized text-to-image generation enables users to create custom AI models that generate their unique concepts—specific objects or artistic styles—achieving unprecedented creative control. However, deploying a large repository of personalized checkpoints faces two critical challenges: (1) ambiguous user prompts make it difficult to match the intended checkpoint in large repositories, and (2) standard post-training quantization methods degrade personalized diffusion checkpoints’ image quality. We analyze the importance of reasoning over checkpoint metadata and clarifying user prompts for intent-aligned checkpoint selection. Additionally, we find that trigger tokens for personalized diffusion play a crucial role in quantization. To address the challenges, we propose PersonalQ, a unified system with two components: Check-in analyzes checkpoint repositories and clarifies user intent for intent-aligned selection, and TAQ (Trigger-Aware Quantization), which protects the trigger-token-related representation to deliver high-quality inference from the chosen checkpoint under quantization. On our Repo-Prompts benchmark, PersonalQ achieves an 89% checkpoint-selection preference win rate and a 4.42/5 intent score. Across benchmarks, TAQ reduces inference memory by up to 75% while maintaining strong text-image alignment (CLIP score 0.297 vs. 0.315 at full precision) and image fidelity (FID 11.03 at W8A8 vs. 10.96 at full precision), enabling scalable deployment of personalized models without compromising quality."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Personalized text-to-image generation"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/50f61b6537bdaf1e298c0bcf4390b40ad56a54eb.pdf"
},
"primary_area": {
"value": "applications to computer vision, audio, language, and other modalities"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/4878d33f88b5ea78ce8e4633adfff8251e992811.zip"
},
"title": {
"value": "PersonalQ: Select, Quantize, and Serve Personalized Diffusion Models for Efficient Inference"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "03MfCNn3pF",
"id": "03MfCNn3pF",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission6759/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897895805,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission6759/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission6759/Authors"
]
} | |
2,026 | 03QzvMzxVM | [
2,
4,
4,
4
] | [
{
"content": "This work presents Robust-NLL, which serves as a plug-and-play loss replacing vanilla NLL loss for robust uncertainty-aware training against label-space outliers. The proposed loss function uses softmax reweighting over sample losses to filter out outliers. The author also provides theoretical ana... | {
"cdate": 1758019401870,
"content": {
"TLDR": {
"value": "We introduce Robust-NLL for modeling uncertainty under the presence of outliers."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025robust,\ntitle={Robust Uncertainty-Aware Learning via Boltzmann-weighted {NLL}},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=03QzvMzxVM},\nnote={under review}\n}"
},
"abstract": {
"value": "Uncertainty estimation is critical for deploying deep learning models in high-stakes applications such as autonomy and decision-making. While prior works on data uncertainty modeling estimate aleatoric uncertainty by minimizing the negative log-likelihood (NLL) loss, they often fail under the presence of outliers. To address this limitation, we introduce Robust-NLL, a drop-in replacement for vanilla NLL that filters noisy or adversarial samples. Robust-NLL learns robust uncertainty estimates in neural networks through a Boltzmann-weighted NLL loss that requires no architectural changes, additional parameters, or iterative procedures, and acts as a plug-and-play loss function that maintains full differentiability and mini-batch compatibility. We evaluate our approach on synthetic regression tasks and real-world visual localization benchmarks with injected outliers. Experimental results demonstrate that simply replacing NLL with Robust-NLL consistently improves both prediction accuracy and reliability of uncertainty estimates, achieving substantial performance gains across diverse tasks and architectures."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"robust estimation",
"uncertainty estimation"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/444e8304cd012c1ab5fb9f3ae96a85fe575c79e2.pdf"
},
"primary_area": {
"value": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Robust Uncertainty-Aware Learning via Boltzmann-weighted NLL"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "03QzvMzxVM",
"id": "03QzvMzxVM",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission7389/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897855752,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission7389/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission7389/Authors"
]
} | |
2,026 | 03ccrSpjOx | [
4,
4,
4,
6
] | [
{
"content": "The paper studies how deliberation format shapes value expression and consensus in LLM-LLM debates over everyday moral dilemmas. Using 1,000 AITA cases, the authors run pairwise and three-way debates among GPT-4.1, Claude 3.7 Sonnet, and Gemini 2.0 Flash in two settings: synchronous (parallel) and... | {
"cdate": 1758148909076,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025deliberative,\ntitle={Deliberative Dynamics and Value Alignment in {LLM} Debates},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=03ccrSpjOx},\nnote={under review}\n}"
},
"abstract": {
"value": "As large language models (LLMs) are increasingly deployed in sensitive everyday contexts -- offering personal advice, mental health support, and moral guidance -- understanding their elicited values in navigating complex moral reasoning is essential. Most evaluations study this sociotechnical alignment through single-turn prompts, but it is unclear if these findings extend to multi-turn settings where values emerge through dialogue, revision, and consensus. We address this gap using LLM debate to examine deliberative dynamics and value alignment in multi-turn settings by prompting subsets of three models (GPT-4.1, Claude 3.7 Sonnet, and Gemini 2.0 Flash) to collectively assign blame in 1,000 everyday dilemmas from Reddit's \"Am I the Asshole\" community. We use both synchronous (parallel responses) and round-robin (sequential responses) formats to test order effects and verdict revision. Our findings show striking behavioral differences. In the synchronous setting, GPT showed strong inertia (0.6-3.1% revision rates) while Claude and Gemini were far more flexible (28-41%). Value patterns also diverged: GPT emphasized personal autonomy and direct communication, while Claude and Gemini prioritized empathetic dialogue. Certain values proved especially effective at driving verdict changes. We further find that deliberation format had a strong impact on model behavior: GPT and Gemini stood out as highly conforming relative to Claude, with their verdict behavior strongly shaped by order effects. These results show how deliberation format and model-specific behaviors shape moral reasoning in multi-turn interactions, underscoring that sociotechnical alignment depends on how systems structure dialogue as much as on their outputs."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"sociotechnical alignment",
"multi-agent debate",
"multi-turn interaction"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/53b15162b8d0641d663ed2799ca10373fb23b76b.pdf"
},
"primary_area": {
"value": "alignment, fairness, safety, privacy, and societal considerations"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Deliberative Dynamics and Value Alignment in LLM Debates"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "03ccrSpjOx",
"id": "03ccrSpjOx",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission9918/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897686075,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission9918/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission9918/Authors"
]
} | |
2,026 | 03fFxN6Orj | [
4,
2,
4
] | [
{
"content": "This paper proposed the Adviser-Actor-Critic (AAC) framework, targeting steady-state error reduction for high-precision robotic control tasks in reinforcement learning. AAC augments standard actor-critic architectures with an additional “adviser” module, implemented as a PI controller, that genera... | {
"cdate": 1758271601146,
"content": {
"TLDR": {
"value": "Adviser-Actor-Critic (AAC) combines reinforcement learning with a novel adviser to generate virtual goals, effectively reducing steady-state errors by over 80% in high-precision robotic control tasks."
},
"_bibtex": {
"value": "@misc{\nchen2025adviseractorcritic,\ntitle={Adviser-Actor-Critic: Reducing Steady-State Error in Reinforcement Learning for Robotics Control},\nauthor={Donghe Chen and Jiaxuan Yue and Yubin Peng and Tengjie Zheng and Han Wang and Chaoran Qu and Lin Cheng},\nyear={2025},\nurl={https://openreview.net/forum?id=03fFxN6Orj}\n}"
},
"abstract": {
"value": "High-precision control tasks present substantial challenges for reinforcement learning (RL) algorithms, frequently resulting in suboptimal performance attributed to network approximation inaccuracies and inadequate sample quality. While existing RL frameworks can achieve task completion at coarse precision levels, steady-state tracking errors remain a critical limitation that prevents achieving sub-hardware-level precision. We introduce Adviser-Actor-Critic (AAC), designed to address this precision control dilemma by combining the precision of feedback control theory with the adaptive learning capability of RL and featuring an Adviser that mentors the actor to refine control actions, thereby enhancing the precision of goal attainment. Through extensive benchmark environments from gymnasium-robotics, coupled with real-world quadcopter attitude control, AAC significantly outperforms standard RL algorithms in precision-critical tasks while demonstrating an average $>80\\%$ steady-state error reduction compared to baseline methods."
},
"anonymous_url": null,
"authorids": {
"value": [
"~Donghe_Chen1",
"~Jiaxuan_Yue2",
"~Yubin_Peng1",
"~Tengjie_Zheng1",
"~Han_Wang17",
"~Chaoran_Qu1",
"~Lin_Cheng7"
]
},
"authors": {
"value": [
"Donghe Chen",
"Jiaxuan Yue",
"Yubin Peng",
"Tengjie Zheng",
"Han Wang",
"Chaoran Qu",
"Lin Cheng"
]
},
"code_of_ethics": null,
"keywords": {
"value": [
"reinforcement learning",
"robotics",
"control system"
]
},
"no_acknowledgement_section": null,
"paperhash": {
"value": "chen|adviseractorcritic_reducing_steadystate_error_in_reinforcement_learning_for_robotics_control"
},
"pdf": {
"value": "/pdf/635d6df0d70e8cc046d12fa468fe1667715b0a02.pdf"
},
"primary_area": {
"value": "reinforcement learning"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Adviser-Actor-Critic: Reducing Steady-State Error in Reinforcement Learning for Robotics Control"
},
"venue": {
"value": "ICLR 2026 Conference Withdrawn Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Withdrawn_Submission"
}
},
"forum": "03fFxN6Orj",
"id": "03fFxN6Orj",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/-/Withdrawn_Submission"
],
"license": "CC BY 4.0",
"mdate": 1762955287461,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission17048/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission17048/Authors"
]
} | |
2,026 | 03jzVlLxEe | [
6,
6,
4,
4
] | [
{
"content": "The authors propose **NERVE**, a noise- and variability-robust EEG foundation model designed to address key challenges in EEG analysis, including low signal-to-noise ratios (SNR), high inter-sample variability, and spatial dependencies arising from electrode placement in acquisition systems. The p... | {
"cdate": 1758337883115,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025nerve,\ntitle={{NERVE}: Noise-Variability-Robust {EEG} Foundation Model with Electrode-Brain Interactions},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=03jzVlLxEe},\nnote={under review}\n}"
},
"abstract": {
"value": "Electroencephalography (EEG) is an indispensable modality for measuring and recording brain electrical activity, with broad applications in brain–computer interfaces (BCI) and healthcare. While early EEG models predominantly adopted supervised learning methods due to the scarcity of large-scale datasets and the heterogeneity across tasks and datasets, the recent success of large foundation models has driven increasing efforts to build EEG foundation models. However, most existing studies focus on handling signals with varying formats while overlooking inherent characteristics of EEG signals during acquisition, including low signal-to-noise ratios (SNR), high variability across samples, and spatial dependencies arising from electrode placement within the acquisition system. To address these challenges, we propose NERVE, a novel noise-variability-robust EEG foundation model with electrode-brain interactions. Specifically, pre-training of NERVE begins with learning a noise-robust neural tokenizer that encodes EEG patches into discrete neural tokens. The tokenizer is trained through denoising temporal–spectral prediction to reconstruct temporal and frequency information of the original signal from noise-augmented inputs. NERVE is further pretrained to predict the neural codes of masked EEG patches, integrated with a variability-robust objective that promotes uniform EEG representations. To incorporate spatial structure in EEG, we propose an electrode-position-aware transformer as the backbone for both the tokenizer and the foundation model. It enables the model to capture spatial dependencies among electrodes and brain regions via attention mechanisms. NERVE demonstrates competitive performance across diverse BCI tasks and improved robustness to noise and variability compared to existing EEG foundation models."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Foundation model",
"Electroencephalography",
"EEG",
"Self-supervised learning",
"Pre-training"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/2af8f2986c76341d381f0b7aced096521dd9722f.pdf"
},
"primary_area": {
"value": "foundation or frontier models, including LLMs"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "NERVE: Noise-Variability-Robust EEG Foundation Model with Electrode-Brain Interactions"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "03jzVlLxEe",
"id": "03jzVlLxEe",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission22991/-/Full_Submission",
"ICLR.cc/2026/Conference/-/Edit"
],
"license": "CC BY 4.0",
"mdate": 1759896837180,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission22991/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission22991/Authors"
]
} | |
2,026 | 03qTI3NKqi | [
4,
4,
4,
4
] | [
{
"content": "This work found that previous soft prompts often disrupted information flow and reduced reasoning. They argue that soft prompts should not be limited to the activation and guidance stages but should be inserted into appropriate stages to ensure smooth information flow between layers. Therefore, th... | {
"cdate": 1758191821554,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025unlocking,\ntitle={Unlocking Coherent Reasoning in {LLM}s with Hierarchical Soft Prompts},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=03qTI3NKqi},\nnote={under review}\n}"
},
"abstract": {
"value": "Large language models (LLMs) exhibit strong reasoning capabilities in complex tasks. Soft prompt tuning, as a lightweight approach, injects trainable vectors into the input to guide the reasoning process and enhance model performance. Prior studies show that soft prompts effectively activate prior knowledge and improve problem understanding in the early stages of reasoning. However, when they continue to exert strong influence in the middle and later stages, they often disrupt the information flow and degrade reasoning performance. Based on this observation, we argue that the role of soft prompts should not be confined to a single stage of activation and guidance. Instead, they should be inserted at appropriate stages to ensure smooth information transmission across layers. Existing methods, however, typically rely on one-shot static injection and cannot dynamically regulate prompts across stages, leading to functional mismatches during reasoning. To address this limitation, we propose a dynamic hierarchy-aware mechanism(DHAM). This mechanism first employs hierarchical clustering to derive stage-specific representations, and then leverages the semantic guidance capability of soft prompts to adaptively align and activate them, ensuring effective coordination across reasoning stages. \nDHAM yields consistent gains across models and benchmarks (e.g., 29.5\\%→43.8\\% on Llama-2-13B/GSM8K), with ablations showing CKA clustering and moderate stage numbers (e.g., $G=3/4$) perform best, consistent with the stable information flow hypothesis."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Large Language Models",
"Complex Reasoning",
"Soft Prompt Tuning"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/511e5f43840e80d2617f1692ac8a2bf18b3b16d7.pdf"
},
"primary_area": {
"value": "foundation or frontier models, including LLMs"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Unlocking Coherent Reasoning in LLMs with Hierarchical Soft Prompts"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "03qTI3NKqi",
"id": "03qTI3NKqi",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission11167/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897603181,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission11167/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission11167/Authors"
]
} | |
2,026 | 03u504EDJp | [
2,
4,
6,
2,
2
] | [
{
"content": "This paper introduces APO, a new framework for distilling reasoning capabilities from multiple MLLMs that exhibit conceptual drift, defined as variability in their reasoning behaviors or conclusions. The core idea is that APO aggregates all available reasoning trajectories and learns to prefer the... | {
"cdate": 1756744193214,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025learning,\ntitle={Learning from All: Concept Alignment for Autonomous Distillation from Multiple Drifting {MLLM}s},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=03u504EDJp},\nnote={under review}\n}"
},
"abstract": {
"value": "This paper identifies a critical yet underexplored challenge in distilling from multi-modal large language models (MLLMs): the reasoning trajectories generated by multiple drifting teachers exhibit concept drift, whereby their reasoning distributions evolve unpredictably and transmit biases to the student model, ultimately compromising its performance. To tackle this issue, we pioneer a theoretical connection between concept drift and knowledge distillation, casting the non-stationary reasoning dynamics from multiple MLLM teachers as next-token prediction of multi-stream reasoning trajectories. Guided by concept drift, we introduce the “learn–compare–critique” paradigm, culminating in autonomous preference optimization (APO). Under the active guidance of the teachers, the student model first learns and self-distils preferred thinking by comparing multiple teachers. It then engages in critical reflection over the drifting inference from teachers, performing concept alignment through APO, ultimately yielding a robust, consistent, and generalizable model. Extensive experiments demonstrate our superior performance of consistency, robustness and generalization within knowledge distillation. Besides, we also contributed a large-scale dataset CXR-MAX (Multi-teachers Alignment X-rays), comprising 170,982 distilled reasoning trajectories derived from publicly accessible MLLMs based on MIMIC-CXR. Our code and data are public at: https://anonymous.4open.science/r/Autonomous-Distillation/."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"concept drift",
"transfer learning",
"multi view",
"knowledge distillation",
"multi modal large language model"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/fe4866ea94ed809fb98d3d8b49b15b242306766f.pdf"
},
"primary_area": {
"value": "learning theory"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/d3f2bf191b959b040fec6edae75de60b04403059.pdf"
},
"title": {
"value": "Learning from All: Concept Alignment for Autonomous Distillation from Multiple Drifting MLLMs"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "03u504EDJp",
"id": "03u504EDJp",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission525/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759898255701,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission525/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission525/Authors"
]
} | |
2,026 | 040ClRXMf3 | [
6,
8,
2,
8
] | [
{
"content": "This paper proposes a new algorithm to extract cardinal-minimal sufficient explanations for Neural Additive Models (NAMs).\nIt does so by exploiting key design choices of NAMs, showing how this family of models supports explanations with guarantees.\n\nThis is achieved as follows. First, the paper... | {
"cdate": 1758298867680,
"content": {
"TLDR": {
"value": "Our approach constructs provably sufficient and (globally) cardinal-minimal explanations for neural additive models with improved runtime complexity."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025provably,\ntitle={Provably Explaining Neural Additive Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=040ClRXMf3},\nnote={under review}\n}"
},
"abstract": {
"value": "Despite significant progress in post-hoc explanation methods for neural networks, many remain heuristic and lack provable guarantees. A key approach for obtaining explanations with provable guarantees is by identifying a *(globally) cardinal-minimal* subset of input features which by itself is *provably sufficient* to determine the prediction. However, for standard neural networks, this task is often computationally infeasible, as it demands a worst-case *exponential* number of verification queries in the number of input features, each of which is NP-hard. In this work, we show that for Neural Additive Models (NAMs), a recent and more interpretable neural network family, we can *efficiently* generate explanations with such guarantees. We present a new model-specific algorithm for NAMs that generates provably (globally) cardinal-minimal explanations using only a *logarithmic* number of verification queries in the number of input features, after a parallelized preprocessing step with logarithmic runtime in the required precision is applied to each small univariate NAM component. Our algorithm not only makes the task of obtaining (globally) cardinal minimal explanations feasible, but even outperforms existing algorithms designed to find *(locally) subset-minimal* explanations -- which may be larger and less informative but easier to compute -- despite our algorithm solving a much more difficult task. Our experiments demonstrate that, compared to previous algorithms, our approach provides provably smaller explanations than existing works and substantially reduces the computation time. Moreover, we show that our generated provable explanations offer benefits that are unattainable by standard sampling-based techniques typically used to interpret NAMs."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"explainability",
"XAI",
"explainable AI",
"formal verification",
"sufficient explanations"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/d5a73d9cf5e02a90d26e33e9057769ff66ff64fa.pdf"
},
"primary_area": {
"value": "interpretability and explainable AI"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/688a5ff66ccb15d28a06f568b0f04b60f4413e61.zip"
},
"title": {
"value": "Provably Explaining Neural Additive Models"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "040ClRXMf3",
"id": "040ClRXMf3",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission19723/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897022892,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission19723/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission19723/Authors"
]
} | |
2,026 | 04HwYGgp2w | [
6,
8,
6,
6
] | [
{
"content": "In this paper,the authors introduces ImageDoctor, a unified,multi-aspect evaluation framework for Text-to Image(T2I) models. Unlike previous methods that provide a single scalar, ImageDoctor assesses image quality across four dimensions: plausibility, semantic alignment, aesthetics, and overall qu... | {
"cdate": 1757544654492,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025imagedoctor,\ntitle={ImageDoctor: Diagnosing Text-to-Image Generation via Grounded Image Reasoning},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=04HwYGgp2w},\nnote={under review}\n}"
},
"abstract": {
"value": "The rapid advancement of text-to-image (T2I) models has increased the need for reliable human preference modeling, a demand further amplified by recent progress in reinforcement learning for preference alignment. However, existing approaches typically quantify the quality of a generated image using a single scalar, limiting their ability to provide comprehensive and interpretable feedback on image quality. To address this, we introduce ImageDoctor, a unified multi-aspect T2I model evaluation framework that assesses image quality across four complementary dimensions: plausibility, semantic alignment, aesthetics, and overall quality. ImageDoctor also provides pixel-level flaw indicators in the form of heatmaps, which highlight misaligned or implausible regions, and can be used as a dense reward for T2I model preference alignment. Inspired by the diagnostic process, we improve the detail sensitivity and reasoning capability of ImageDoctor by introducing a ``look-think-predict\" paradigm, where the model first localizes potential flaws, then generates reasoning, and finally concludes the evaluation with quantitative scores. Built on top of a vision-language model and trained through a combination of supervised fine-tuning and reinforcement learning, ImageDoctor demonstrates strong alignment with human preference across multiple datasets, establishing its effectiveness as an evaluation metric. Furthermore, when used as a reward model for preference tuning, ImageDoctor significantly improves generation quality—achieving an improvement of 10% over scalar-based reward models."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Image reward model"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/ab62de115d368d82b0351f14bb9466e9bbe97c92.pdf"
},
"primary_area": {
"value": "applications to computer vision, audio, language, and other modalities"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "ImageDoctor: Diagnosing Text-to-Image Generation via Grounded Image Reasoning"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "04HwYGgp2w",
"id": "04HwYGgp2w",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission3835/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759898067519,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission3835/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission3835/Authors"
]
} | |
2,026 | 04JkPDiCnp | [
2,
6,
4,
2
] | [
{
"content": "This paper introduces InternAgent-DR, a multi-agent deep-research framework that models scientific reasoning as a dynamic structured knowledge flow. Instead of relying on a linear task sequence, InternAgent-DR represents research workflows as directed acyclic graphs whose nodes correspond to subta... | {
"cdate": 1756820032542,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025internagentdr,\ntitle={InternAgent-{DR}: Advancing deep research with dynamic structured knowledge flow},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=04JkPDiCnp},\nnote={under review}\n}"
},
"abstract": {
"value": "Deep research is an inherently challenging task that demands both breadth and depth of thinking. It involves navigating diverse knowledge spaces and reasoning over complex, multi-step dependencies, which presents substantial challenges for agentic systems. To address this, we propose InternAgent-DR (Deep Research), a multi-agent framework that actively constructs and evolves a dynamic structured knowledge flow to drive subtask execution and reasoning. InternAgent-DR is capable of strategically planning and expanding the knowledge flow to enable parallel exploration and hierarchical task decomposition, while also adjusting the knowledge flow in real time based on feedback from intermediate reasoning outcomes and insights. InternAgent-DR achieves state-of-the-art performance on both general and scientific benchmarks, including GAIA, HLE, GPQA and TRQA, demonstrating its effectiveness in multi-disciplinary research scenarios and its potential to advance scientific discovery."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"deep research",
"multi-agent",
"reasoning model"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/1733de55f54fb9280e4bfee98aaf47ded2d07fd1.pdf"
},
"primary_area": {
"value": "applications to computer vision, audio, language, and other modalities"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "InternAgent-DR: Advancing deep research with dynamic structured knowledge flow"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "04JkPDiCnp",
"id": "04JkPDiCnp",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission830/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759898239693,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission830/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission830/Authors"
]
} | |
2,026 | 04Tfwy3LLC | [
2,
6,
4,
8
] | [
{
"content": "The paper relates to the pruning of LLM layers. The paper consists of three main parts:\n1. Discussion of criteria for identifying prunable layers\n2. Comparison between LoRA and partial fine-tuning methods for recovering accuracy after pruning\n3. Theoretical analysis of gradient flow in the pres... | {
"cdate": 1757254648198,
"content": {
"TLDR": {
"value": "This paper presents a theoretical and empirical analysis of layer pruning in Large Language Models, aiming to improve and refine pruning strategies."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025reassessing,\ntitle={Reassessing Layer Pruning in {LLM}s: New Insights and Methods},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=04Tfwy3LLC},\nnote={under review}\n}"
},
"abstract": {
"value": "Although large language models (LLMs) have achieved remarkable success across various domains, their considerable scale necessitates substantial computational resources, posing significant challenges for deployment in resource-constrained environments. Layer pruning, as a simple yet effective compression method, removes layers of a model directly, reducing computational overhead. However, what are the best practices for layer pruning in LLMs? Are sophisticated layer selection metrics truly effective? Does the LoRA (Low-Rank Approximation) family, widely regarded as a leading method for pruned model fine-tuning, truly meet expectations when applied to post-pruning fine-tuning? To answer these questions, we dedicate thousands of GPU hours to benchmarking layer pruning in LLMs and gaining insights across multiple dimensions. Our results demonstrate that a simple approach, i.e., pruning the final layers followed by fine-tuning the lm\\_head and the remaining last three layers, yields remarkably strong performance. These pruning strategies are further supported by theoretical analyses based on the gradient flow. Following this guide, our method surpasses existing state-of-the-art pruning methods by $5.62\\%$–$17.27\\%$ on Llama-3.1-8B-It, by $2.36\\%$–$19.45\\%$ on Llama-3-8B and by $4.34\\%$–$9.59\\%$ on Llama-3-70B. The code is available on GitHub."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Large Language Model",
"Layer Pruning",
"Model Compression"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/c6ed1e0f689d0744c27ac966827d51d77a626dce.pdf"
},
"primary_area": {
"value": "foundation or frontier models, including LLMs"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Reassessing Layer Pruning in LLMs: New Insights and Methods"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "04Tfwy3LLC",
"id": "04Tfwy3LLC",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759898126388,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission2804/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission2804/Authors"
]
} | |
2,026 | 04h40hEgTj | [
6,
6,
2,
4
] | [
{
"content": "In this paper, the authors aimed at creating a family of toy models for exploring the known challenge of long-context learning for LLM. The proposed toy model have different time series data interleaved with distinct labels. The authors found that LLM developed two distinct learning mechanisms in ... | {
"cdate": 1758340263445,
"content": {
"TLDR": {
"value": "We introduce a new family of toy problems that combine features of linear-regression-style continuous in-context learning (ICL) with discrete associative recall and find distinct learning dynamics for different prediction mechanisms."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025decomposing,\ntitle={Decomposing Prediction Mechanisms for In-context Recall},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=04h40hEgTj},\nnote={under review}\n}"
},
"abstract": {
"value": "We introduce a new family of toy problems to explore challenges with long context learning and associative recall in transformer models. Our setup involves interleaved segments of observations from randomly drawn linear deterministic dynamical systems. Each system is associated with a discrete symbolic label that must be learned in-context since these associations randomly shuffle between training instances.\n\nVia out-of-distribution experiments we find that learned next-token prediction for this toy problem involves at least two separate mechanisms. One \"label-based\" mechanism uses the discrete symbolic labels to do the associative recall required to predict the start of a resumption of a previously seen system's observations. The second ``observation-based'' mechanism largely ignores the discrete symbolic labels and performs a prediction based on the state observations previously seen in context. These two mechanisms have different learning dynamics: the second mechanism develops much earlier than the first.\n\nThe behavior of our toy model suggested concrete experiments that we performed with OLMo training checkpoints on an ICL translation task. We see a similar phenomenon: the model learns to continue a translation task in-context earlier than it decisively learns to in-context identify the meaning of a symbolic label telling it to translate."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"emergence",
"in-context learning",
"time-series",
"associative recall",
"learning dynamics"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/874dd26fa4acf6f26e690461d6232071b158fd84.pdf"
},
"primary_area": {
"value": "interpretability and explainable AI"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Decomposing Prediction Mechanisms for In-context Recall"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "04h40hEgTj",
"id": "04h40hEgTj",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission23149/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759896830101,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission23149/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission23149/Authors"
]
} | |
2,026 | 053vZMxDB5 | [
2,
8,
4
] | [
{
"content": "This paper presents a reinforcement learning (RL) approach for learning from signal temporal logic (STL) to make learning more feasible for long-horizon tasks. The novel model-free approach divides and flattens complex STL formulas and searches for time-variable actualizations via Metropolis-Hasti... | {
"cdate": 1756884774931,
"content": {
"TLDR": {
"value": "We design a Reinforcement Learning framework based on time variables and task decomposition to solve Signal Temporal Logic tasks"
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025tgpo,\ntitle={{TGPO}: Temporal Grounded Policy Optimization for Signal Temporal Logic Tasks},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=053vZMxDB5},\nnote={under review}\n}"
},
"abstract": {
"value": "Learning control policies for complex, long-horizon tasks is a central challenge in robotics and autonomous systems. Signal Temporal Logic (STL) offers a powerful and expressive language for specifying such tasks, but its non-Markovian nature and inherent sparse reward make it difficult to be solved via standard Reinforcement Learning (RL) algorithms. Prior RL approaches focus only on limited STL fragments or use STL robustness scores as sparse terminal rewards. In this paper, we propose TGPO, Temporal Grounded Policy Optimization, to solve general STL tasks. TGPO decomposes STL into timed subgoals and invariant constraints and provides a hierarchical framework to tackle the problem. The high-level component of TGPO proposes concrete time allocations for these subgoals, and the low-level time-conditioned policy learns to achieve the sequenced subgoals using a dense, stage-wise reward signal. During inference, we sample various time allocations and select the most promising assignment for the policy network to rollout the solution trajectory. To foster efficient policy learning for complex STL with multiple subgoals, we leverage the learned critic to guide the high-level temporal search via Metropolis-Hastings sampling, focusing exploration on temporally feasible solutions. We conduct experiments on five environments, ranging from low-dimensional navigation to manipulation, drone, and quadrupedal locomotion. Under a wide range of STL tasks, TGPO significantly outperforms state-of-the-art baselines (especially for high-dimensional and long-horizon cases), with an average of 31.6% improvement in task success rate compared to the best baseline."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Reinforcement Learning; Signal Temporal Logic"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/524211ceccea6ca532fc8ec47c9c896c13dd9fa7.pdf"
},
"primary_area": {
"value": "applications to robotics, autonomy, planning"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "TGPO: Temporal Grounded Policy Optimization for Signal Temporal Logic Tasks"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "053vZMxDB5",
"id": "053vZMxDB5",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission1461/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759898207954,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission1461/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission1461/Authors"
]
} | |
2,026 | 05NHmcEpNk | [
8,
4,
8
] | [
{
"content": "This paper introduces CT-MLE, a model-based algorithm for continuous-time reinforcement learning (CTRL) that uses maximum likelihood estimation (MLE) of the state marginal density instead of directly modeling system dynamics.\nThe key idea is to achieve instance-dependent adaptivity, where the alg... | {
"cdate": 1758213925539,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025instancedependent,\ntitle={Instance-Dependent Continuous-Time Reinforcement Learning via Maximum Likelihood Estimation},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=05NHmcEpNk},\nnote={under review}\n}"
},
"abstract": {
"value": "Continuous-time reinforcement learning (CTRL) provides a natural framework for sequential decision-making in dynamic environments where interactions evolve continuously over time. While CTRL has shown growing empirical success, its ability to adapt to varying levels of problem difficulty remains poorly understood. In this work, we investigate the instance-dependent behavior of CTRL and introduce a simple, model-based algorithm built on maximum likelihood estimation (MLE) with a general function approximator. Unlike existing approaches that estimate system dynamics directly, our method estimates the state marginal density to guide learning. We establish instance-dependent performance guarantees by deriving a regret bound that scales with the total reward variance and measurement resolution. Notably, the regret becomes independent of the specific measurement strategy when the observation frequency adapts appropriately to the problem’s complexity. To further improve performance, our algorithm incorporates a randomized measurement schedule that enhances sample efficiency without increasing measurement cost. These results highlight a new direction for designing CTRL algorithms that automatically adjust their learning behavior based on the underlying difficulty of the environment."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Continuous-time reinforcement learning"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/9f4ff6eac9d7af34e021903665ab4988e2f46ad6.pdf"
},
"primary_area": {
"value": "learning theory"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Instance-Dependent Continuous-Time Reinforcement Learning via Maximum Likelihood Estimation"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "05NHmcEpNk",
"id": "05NHmcEpNk",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission13133/-/Full_Submission",
"ICLR.cc/2026/Conference/Submission13133/-/Rebuttal_Revision"
],
"license": "CC BY 4.0",
"mdate": 1763388667273,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission13133/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission13133/Authors"
]
} | |
2,026 | 05PqjBzN6S | [
4,
2,
6
] | [
{
"content": "This paper addresses the problem of determining when sufficient data is available to safely retrain a model after a sudden concept drift. The authors propose CALIPER, a model-agnostic and data-only test to estimate this required post-drift data size. The core idea is grounded in the concept of \"s... | {
"cdate": 1758350444098,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025when,\ntitle={When to Retrain after Drift: A Data-Only Test of Post-Drift Data Size Sufficiency},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=05PqjBzN6S},\nnote={under review}\n}"
},
"abstract": {
"value": "Sudden concept drift makes previously trained predictors unreliable, yet deciding when to retrain and what post-drift data size is sufficient is rarely addressed. We propose CALIPER —a detector- and model-agnostic, data-only test that estimates the post-drift data size required for stable retraining. CALIPER exploits state dependence in streams generated by dynamical systems: we run a single-pass weighted local regression over the post-drift window and track a one-step proxy error as a function of a locality parameter $\\theta$. When an effective sample size gate is satisfied, a monotonically non-increasing trend in this error with increasing a locality parameter indicates that the data size is sufficiently informative for retraining.\nWe also provide a theoretical analysis of our CALIPER, and we show that the algorithm has a low per-update time and memory. Across datasets from four heterogeneous domains, three learner families, and two detectors, CALIPER consistently matches or exceeds the best fixed data size for retraining while incurring negligible overhead and often outperforming incremental updates. CALIPER closes the gap between drift detection and data-sufficient adaptation in streaming learning."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Concept drift",
"Stream learning",
"Data sufficiency",
"Time series"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/c36ecf51e14859470565e33d2e39e69232a4cb26.pdf"
},
"primary_area": {
"value": "learning on time series and dynamical systems"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "When to Retrain after Drift: A Data-Only Test of Post-Drift Data Size Sufficiency"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "05PqjBzN6S",
"id": "05PqjBzN6S",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission23926/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759896790097,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission23926/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission23926/Authors"
]
} | |
2,026 | 05SHW9ai9e | [
4,
2,
4,
4
] | [
{
"content": "To address DocQA limitations (single-modality bias, isolated RAG, long-document overload), this paper proposes MDocAgent—a framework integrating dual RAG (text via ColBERTv2, image via ColPali) and 5 collaborative agents (General, Critical, Text, Image, Summarizing). Evaluated on 5 benchmarks (MML... | {
"cdate": 1758214136657,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025mdocagent,\ntitle={{MD}ocAgent: A Multi-Modal Multi-Agent Framework for Document Question Answering},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=05SHW9ai9e},\nnote={under review}\n}"
},
"abstract": {
"value": "Document Question Answering (DocQA) is a very common task. Existing methods using Large Language Models (LLMs) or Large Vision Language Models (LVLMs) and Retrieval Augmented Generation (RAG) often prioritize information from a single modal, failing to effectively integrate textual and visual cues. These approaches struggle with complex multi-modal reasoning, limiting their performance on real-world documents. We present MDocAgent (A Multi-Modal Multi-Agent Framework for Document Question Answering), a novel RAG and multi-agent framework that leverages both text and image. Our system employs five specialized agents: a general agent, a critical agent, a text agent, an image agent and a summarizing agent. These agents engage in multi-modal context retrieval, combining their individual insights to achieve a more comprehensive understanding of the document's content. This collaborative approach enables the system to synthesize information from both textual and visual components, leading to improved accuracy in question answering. Preliminary experiments on five benchmarks like MMLongBench, LongDocURL demonstrate the effectiveness of our MDocAgent, achieve an average improvement of 12.1% compared to current state-of-the-art method. This work contributes to the development of more robust and comprehensive DocQA systems capable of handling the complexities of real-world documents containing rich textual and visual information."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Multimodal",
"DocQA",
"RAG",
"LVLM"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/2ddcd015efb50efa2aa66b781add39ffb4dc6e92.pdf"
},
"primary_area": {
"value": "applications to computer vision, audio, language, and other modalities"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "MDocAgent: A Multi-Modal Multi-Agent Framework for Document Question Answering"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "05SHW9ai9e",
"id": "05SHW9ai9e",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission13150/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897460751,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission13150/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission13150/Authors"
]
} | |
2,026 | 05THHF0w3y | [
0,
2,
4,
4
] | [
{
"content": "The paper proposes a new method for LLM reasoning, R-Capsule, where LLMs first output high-level plans which are in a latent space and then textual detailed steps and finally the answer. The authors choose several benchmarks on math reasoning (such as GSM-8k) and commensense reasoning (such as str... | {
"cdate": 1757406324840,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025rcapsule,\ntitle={R-Capsule: Compressing High-Level Plans for Efficient Large Language Model Reasoning},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=05THHF0w3y},\nnote={under review}\n}"
},
"abstract": {
"value": "Chain-of-Thought (CoT) prompting has enabled Large Language Models (LLMs) to tackle complex reasoning tasks by generating explicit step-by-step rationales. However, this verbosity incurs significant computational overhead in terms of latency and memory, and can lead to error propagation over long reasoning chains. We propose the \\textbf{Reasoning Capsule}, a novel framework that captures the efficiency of latent reasoning while retaining the transparency of explicit CoT. Our core idea is to compress the high-level strategic plan of a reasoning process into a compact, low-dimensional latent representation---the Reasoning Capsule---while leaving the low-level execution steps explicit. This hybrid approach is grounded in the Information Bottleneck principle, where we learn a capsule that is a \\emph{minimal sufficient statistic} for the reasoning task. Minimality is enforced structurally via a low-dimensional bottleneck, ensuring efficiency. Sufficiency is enforced via a dual-objective function: a primary task loss for answer accuracy and an auxiliary reconstruction loss that ensures the capsule faithfully represents the original textual plan. This reconstruction objective grounds the latent space, making the compressed plan interpretable and robust against uninformative shortcuts. Our framework unifies efficiency, accuracy, and interpretability, significantly reducing the token footprint of reasoning while maintaining or improving performance on complex reasoning benchmarks."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Large Language Model",
"latent reasoning"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/41ed6938581c932dbdf98a17f0863c19cb7cfbde.pdf"
},
"primary_area": {
"value": "foundation or frontier models, including LLMs"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "R-Capsule: Compressing High-Level Plans for Efficient Large Language Model Reasoning"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "05THHF0w3y",
"id": "05THHF0w3y",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission3349/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759898094406,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission3349/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission3349/Authors"
]
} | |
2,026 | 05hNleYOcG | [
2,
4,
2,
2
] | [
{
"content": "The paper introduces PLAGUE, a plug-and-play framework for designing multi-turn jailbreak attacks on large language models (LLMs). Inspired by lifelong-learning and agentic architectures, PLAGUE divides the attack process into three stages — Planner, Primer, and Finisher — enabling adaptable and m... | {
"cdate": 1758135059535,
"content": {
"TLDR": {
"value": "Agentic framework for discovering novel potent multi-turn jailbreak attacks that achieve an attack success rate of 67.3% on Claude Opus 4.1"
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025plague,\ntitle={{PLAGUE}: Plug-and-play Framework for Lifelong Adaptive Generation of Multi-turn Exploits},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=05hNleYOcG},\nnote={under review}\n}"
},
"abstract": {
"value": "Large Language Models (LLMs) are improving at an exceptional rate. With the advent of agentic workflows, multi-turn dialogue has become the de facto mode of interaction with LLMs for completing long and complex tasks. While LLM capabilities continue to improve, they remain increasingly susceptible to jailbreaking, especially in multi-turn scenarios where harmful intent can be subtly injected across the conversation to produce nefarious outcomes. While single-turn attacks have been extensively explored, adaptability, efficiency and effectiveness continue to remain key challenges for their multi-turn counterparts. To address these gaps, we present PLAGUE, a novel plug-and-play framework for designing multi-turn attacks inspired by lifelong-learning agents. PLAGUE dissects the lifetime of a multi-turn attack into three carefully designed phases (Primer, Planner and Finisher) that enable a systematic and information-rich exploration of the multi-turn attack family. Evaluations show that red-teaming agents designed using PLAGUE achieve state-of-the-art jailbreaking results, improving attack success rates (ASR) by more than 30% across leading models in a lesser or comparable query budget. Particularly, PLAGUE enables an ASR (based on StrongReject) of 81.4% on OpenAI's o3 and 67.3% on Claude's Opus 4.1, two models that are considered highly resistant to jailbreaks in safety literature. Our work offers tools and insights to understand the importance of plan initialization, context optimization, and lifelong learning in crafting multi-turn attacks for a comprehensive model vulnerability evaluation."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"LLM Red-Teaming",
"Agentic AI"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/de8dc0979b8266f26b81ee913344d9abba387bb0.pdf"
},
"primary_area": {
"value": "alignment, fairness, safety, privacy, and societal considerations"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/da1b9d173949372d38df20cfd54baf183ccdf1be.zip"
},
"title": {
"value": "PLAGUE: Plug-and-play Framework for Lifelong Adaptive Generation of Multi-turn Exploits"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "05hNleYOcG",
"id": "05hNleYOcG",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission9695/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897703848,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission9695/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission9695/Authors"
]
} | |
2,026 | 05pfP2khzx | [
2,
2,
4
] | [
{
"content": "This paper introduces VIDEOREPAIR, a video refinement framework to correct text-video misalignments. It has three steps: 1. detect misalignment. Finding the issue and region with MLLM. 2. Plan the refinement including preserve the correct parts and construct prompts that could be used to re-genera... | {
"cdate": 1758222291968,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@misc{\nlee2025selfcorrecting,\ntitle={Self-Correcting Text-to-Video Generation with Misalignment Detection and Localized Refinement},\nauthor={Daeun Lee and Jaehong Yoon and Jaemin Cho and Mohit Bansal},\nyear={2025},\nurl={https://openreview.net/forum?id=05pfP2khzx}\n}"
},
"abstract": {
"value": "Recent text-to-video (T2V) diffusion models have made remarkable progress in\ngenerating high-quality and diverse videos. However, they often struggle to align\nwith complex text prompts, particularly when multiple objects, attributes, or spatial\nrelations are specified. We introduce VideoRepair, the first self-correcting,\ntraining-free, and model-agnostic video refinement framework that automatically\ndetects fine-grained text–video misalignments and performs targeted, localized\ncorrections. Our key insight is that even misaligned videos usually contain correctly\nrendered regions that should be preserved rather than regenerated. Building on this\nobservation, VideoRepair proposes a novel region-preserving refinement strategy\nwith three stages: (i) misalignment detection, where systematic MLLM-based evaluation\nwith automatically generated spatio-temporal questions identifies faithful\nand misaligned regions; (ii) refinement planning, which preserves correctly generated\nentities, segments their regions across frames, and constructs targeted prompts\nfor misaligned areas; and (iii) localized refinement, which selectively regenerates\nproblematic regions while preserving faithful content through joint optimization\nof preserved and newly generated areas. This self-correcting, region-preserving\nstrategy converts evaluation signals into actionable guidance for refinement, enabling\nefficient and interpretable corrections. On two challenging benchmarks,\nEvalCrafter and T2V-CompBench, VideoRepair achieves substantial improvements\nover recent baselines across diverse alignment metrics. Comprehensive\nablations further demonstrate the efficiency, robustness, and interpretability of our\nframework."
},
"anonymous_url": null,
"authorids": {
"value": [
"~Daeun_Lee2",
"~Jaehong_Yoon1",
"~Jaemin_Cho1",
"~Mohit_Bansal2"
]
},
"authors": {
"value": [
"Daeun Lee",
"Jaehong Yoon",
"Jaemin Cho",
"Mohit Bansal"
]
},
"code_of_ethics": null,
"keywords": {
"value": [
"Video Generation",
"Multi-agent"
]
},
"no_acknowledgement_section": null,
"paperhash": {
"value": "lee|selfcorrecting_texttovideo_generation_with_misalignment_detection_and_localized_refinement"
},
"pdf": {
"value": "/pdf/92074a4083fee85665efd54a5e543a7af3d7095e.pdf"
},
"primary_area": {
"value": "applications to computer vision, audio, language, and other modalities"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/d49f3262bd569432cfbb01e316e81fba9e473798.zip"
},
"title": {
"value": "Self-Correcting Text-to-Video Generation with Misalignment Detection and Localized Refinement"
},
"venue": {
"value": "ICLR 2026 Conference Withdrawn Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Withdrawn_Submission"
}
},
"forum": "05pfP2khzx",
"id": "05pfP2khzx",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission13771/-/Full_Submission",
"ICLR.cc/2026/Conference/-/Withdrawn_Submission"
],
"license": "CC BY 4.0",
"mdate": 1762964082540,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission13771/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission13771/Authors"
]
} | |
2,026 | 05uq3XUJaT | [
2,
2,
4
] | [
{
"content": "This paper introduces a listwise fine-tuning method for LLM-based text reranking. The method improves three limitations of existing LLM rankers (single-token compression, shallow scoring heads, and pairwise objectives).",
"id": "DvaKUEhgPp",
"rating": 2
},
{
"content": "This paper ... | {
"cdate": 1757411444566,
"content": {
"TLDR": {
"value": "We propose a method to improve the fine-tuning performance of text ranking models by leveraging feature fusion, incorporating customized MLP modules, and optimizing with a listwise loss."
},
"_bibtex": {
"value": "@misc{\nsong2025finetuning,\ntitle={Fine-tuning large language models for text ranking with listwise constraints},\nauthor={Jiawen Song and Bingfei Zhang and Sai Gao and Xueyao Zhang and Wenqing Xu and Guanyu Chen and Junwei Xing and Hui Li and Yunpeng Peng and Zhi Zang},\nyear={2025},\nurl={https://openreview.net/forum?id=05uq3XUJaT}\n}"
},
"abstract": {
"value": "With the rapid adoption of large language models (LLMs) across diverse applications, retrieval augmentation has become a key factor for improving downstream performance. Recent advances show that LLM-based retrieval can substantially enhance ranking quality. In this work, we present a novel LLM-based retrieval framework optimized along three complementary dimensions: (1) a customized attention-based fusion of hidden-layer representations, (2) a dedicated multi-layer perceptron (MLP) module for enriched feature transformation, and (3) a new list-wise learning objective, ListRank loss, to capture fine-grained relevance order. Experimental results demonstrate that our model achieves state-of-the-art performance. The model is publicly available for download on HuggingFace."
},
"anonymous_url": null,
"authorids": {
"value": [
"~Jiawen_Song1",
"~Bingfei_Zhang1",
"~Sai_Gao1",
"~Xueyao_Zhang2",
"~Wenqing_Xu3",
"~Guanyu_Chen14",
"~Junwei_Xing1",
"~Hui_Li58",
"~Yunpeng_Peng2",
"~Zhi_Zang1"
]
},
"authors": {
"value": [
"Jiawen Song",
"Bingfei Zhang",
"Sai Gao",
"Xueyao Zhang",
"Wenqing Xu",
"Guanyu Chen",
"Junwei Xing",
"Hui Li",
"Yunpeng Peng",
"Zhi Zang"
]
},
"code_of_ethics": null,
"keywords": {
"value": [
"Feature fusion",
"listwise",
"LLM",
"rank"
]
},
"no_acknowledgement_section": null,
"paperhash": {
"value": "song|finetuning_large_language_models_for_text_ranking_with_listwise_constraints"
},
"pdf": {
"value": "/pdf/438531bfdc6d7eff6df3c9f4faf576cb9faa1f30.pdf"
},
"primary_area": {
"value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Fine-tuning large language models for text ranking with listwise constraints"
},
"venue": {
"value": "ICLR 2026 Conference Withdrawn Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Withdrawn_Submission"
}
},
"forum": "05uq3XUJaT",
"id": "05uq3XUJaT",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Edit",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission3367/-/Full_Submission",
"ICLR.cc/2026/Conference/-/Withdrawn_Submission"
],
"license": "CC BY 4.0",
"mdate": 1763361432756,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission3367/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission3367/Authors"
]
} | |
2,026 | 0694m9ixnv | [
4,
6,
2
] | [
{
"content": "This paper introduces Instruction Distillation, a new paradigm for improving the quality of low-quality instruction-following data. The authors propose a dataset called MIXTURE that maps multiple low-quality or redundant text inputs to a distilled high-quality target. Building on this dataset, the... | {
"cdate": 1758008662115,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025lmmixup,\ntitle={{LM}-mixup: Text Data Augmentation via Language Model based Mixup},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0694m9ixnv},\nnote={under review}\n}"
},
"abstract": {
"value": "Instruction tuning is crucial for aligning Large Language Models (LLMs), yet the quality of instruction-following data varies significantly. While high-quality data is paramount, it is often scarce; conversely, abundant low-quality data is frequently discarded, leading to substantial information loss. Existing data augmentation methods struggle to augment this low-quality data effectively, and the evaluation of such techniques remains poorly defined. To address this, we formally define the task of *Instruction Distillation*: distilling multiple low-quality and redundant inputs into high-quality and coherent instruction-output pairs. Specifically, we introduce a comprehensive data construction pipeline to create *MIXTURE*, a 144K-sample dataset pairing low-quality or semantically redundant imperfect instruction clusters with their high-quality distillations. We then introduce *LM-Mixup*, by first performing supervised fine-tuning on *MIXTURE* and then optimizing it with reinforcement learning. This process uses three complementary reward signals: quality, semantic alignment, and format compliance, via Group Relative Policy Optimization (GRPO). We demonstrate that *LM-Mixup* effectively augments imperfect datasets: fine-tuning LLMs on its distilled data, which accounts for only about 3% of the entire dataset, not only surpasses full-dataset training but also competes with state-of-the-art high-quality data selection methods across multiple benchmarks. Our work establishes that low-quality data is a valuable resource when properly distilled and augmented with *LM-Mixup*, significantly enhancing the efficiency and performance of instruction-tuned LLMs."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Instruction distillation",
"LM mixup"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/063db25688cafc17b63b0a73cc99a225f64ae83e.pdf"
},
"primary_area": {
"value": "alignment, fairness, safety, privacy, and societal considerations"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "LM-mixup: Text Data Augmentation via Language Model based Mixup"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0694m9ixnv",
"id": "0694m9ixnv",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission7123/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897871663,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission7123/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission7123/Authors"
]
} | |
2,026 | 06I7jcrkW2 | [
6,
6,
4,
8
] | [
{
"content": "This paper tackles the important and challenging problem of accelerating Real-Time TDDFT (RT-TDDFT) computations using deep learning. \nSpecifically, it adopts an autoregressive framework to accelerate the propagations of RT-TDDFT, where the wavefunctions of previous steps are input into the netw... | {
"cdate": 1758291547393,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025orbital,\ntitle={Orbital Transformers for Predicting Wavefunctions in Time-Dependent Density Functional Theory},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=06I7jcrkW2},\nnote={under review}\n}"
},
"abstract": {
"value": "We aim to learn wavefunctions simulated by time-dependent density functional theory (TDDFT), which can be efficiently represented as linear combination coefficients of atomic orbitals. In real-time TDDFT, the electronic wavefunctions of a molecule evolve over time in response to an external excitation, enabling first-principles predictions of physical properties such as optical absorption, electron dynamics, and high-order response. However, conventional real-time TDDFT relies on time-consuming propagation of all occupied states with fine time steps. In this work, we propose OrbEvo, which is based on an equivariant graph transformer architecture and learns to evolve the full electronic wavefunction coefficients across time steps. First, to account for external field, we design an equivariant conditioning to encode both strength and direction of external electric field and break the symmetry from SO(3) to SO(2). Furthermore, we design two OrbEvo models, OrbEvo-FullWF and OrbEvo-DM, using wavefunction pooling and density matrix as interaction method, respectively. Motivated by the central role of the density functional in TDDFT, OrbEvo-DM encodes the density matrix aggregated from all occupied electronic states into feature vectors via tensor contraction, providing a more intuitive approach to learn the time evolution operator. We adopt a training strategy specifically tailored to limit the error accumulation of time-dependent wavefunctions over autoregressive rollout. To evaluate our approach, we generate TDDFT datasets consisting of 5,000 different molecules in the QM9 dataset and 1,500 molecular configurations of the malonaldehyde molecule in the MD17 dataset. Results show that our OrbEvo model accurately captures quantum dynamics of excited states under external field, including time-dependent wavefunctions, time-dependent dipole moment, and optical absorption spectra characterized by dipole oscillator strength. It also shows strong generalization capability on the diverse molecules in the QM9 dataset."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Machine learning density functional theory",
"Time dependent neural PDE solver"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/b9b9470edaaf38e546adf996fb79f0e4341c771e.pdf"
},
"primary_area": {
"value": "applications to physical sciences (physics, chemistry, biology, etc.)"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Orbital Transformers for Predicting Wavefunctions in Time-Dependent Density Functional Theory"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "06I7jcrkW2",
"id": "06I7jcrkW2",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission18854/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897077611,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission18854/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission18854/Authors"
]
} | |
2,026 | 06bDxmgdE0 | [
4,
2,
4
] | [
{
"content": "This paper introduces a novel and highly significant large-scale, multitask benchmark for evaluating speech understanding capabilities across 11 Southeast Asian (SEA) languages. This work directly addresses the critical lack of non-English evaluation frameworks, as current benchmarks are heavily E... | {
"cdate": 1758092746264,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025seaspeechbench,\ntitle={{SEA}-SpeechBench: A Large-Scale Multitask Benchmark for Speech Understanding Across Southeast Asia},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=06bDxmgdE0},\nnote={under review}\n}"
},
"abstract": {
"value": "The rapid advancement of audio and multimodal large language models has unlocked transformative speech understanding capabilities, yet evaluation frameworks remain predominantly English-centric, leaving Southeast Asian (SEA) languages critically underrepresented. We introduce SEA-SpeechBench, the first large-scale multitask benchmark that evaluates speech understanding in 11 SEA languages through more than 97,000 samples and 597 hours of curated audio data. Our benchmark comprises 9 diverse tasks across 3 categories: speech processing (automatic speech recognition, speech translation, spoken question answering), paralinguistic analysis (emotion, gender, age, speaker recognition), and temporal understanding, a novel dimension featuring timestamped content queries and temporal localization within extended audio sequences up to 3 minutes. We implement multilingual prompting in both native SEA languages and English to reflect user interactions with audio-language models. \nEvaluation of leading open-source and proprietary systems reveals marked performance gaps. Across all models, performance remains underwhelming on temporal reasoning, emotion recognition, and speech translation, with most scores falling below 20. Prompting in low-resource languages such as Burmese, Lao, Tamil, and Khmer lag behind English by over 5%.\nOur findings expose critical model limitations and underscore the need for inclusive model development."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Southeast Asian Languages",
"Multilingual Speech Benchmark",
"Audio–language Models"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/5c41f3c5d166884396fa71337e5d8815b0a06417.pdf"
},
"primary_area": {
"value": "datasets and benchmarks"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/3b437ba45f2f0be6c4610c6c40d04907323b6921.pdf"
},
"title": {
"value": "SEA-SpeechBench: A Large-Scale Multitask Benchmark for Speech Understanding Across Southeast Asia"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "06bDxmgdE0",
"id": "06bDxmgdE0",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission8619/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897772905,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission8619/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission8619/Authors"
]
} | |
2,026 | 06upBSlAUy | [
4,
6,
4
] | [
{
"content": "The paper proposes Stabilized and Improved Preference Optimization (SIPO), a framework designed to address two fundamental challenges in applying Direct Preference Optimization (DPO) to diffusion models: training instability and off-policy bias. The authors first conduct a systematic analysis of t... | {
"cdate": 1758341139376,
"content": {
"TLDR": {
"value": "We propose a stabilized and improved preference optimization framework for aligning diffusion generative models with human perferences."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025sipo,\ntitle={{SIPO}: Stabilized and Improved Preference Optimization for Aligning Diffusion Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=06upBSlAUy},\nnote={under review}\n}"
},
"abstract": {
"value": "Preference learning has garnered extensive attention as an effective technique for aligning diffusion models with human preferences in visual generation tasks. However, existing alignment approaches such as Diffusion-DPO suffer from two fundamental challenges: training instability caused by high gradient variances at various timesteps and high parameter sensitivities, and off-policy bias arising from the discrepancy between the optimization data and the policy model's distribution. Our first contribution is a systematical analysis of the diffusion trajectories across different timesteps and identify that the instability primarily originates from early timesteps with low importance weights. To address these issues, we propose SIPO, a Stabilized and Improved preference Optimization framework for aligning diffusion models with human preferences. Concretely, a key gradient, \\emph{i.e.,} DPO-C\\&M is introduced to facilitate stabilize training by clipping and masking uninformative timesteps. Followed by a timestep aware importance re-weighting paradigm to fully correct off-policy bias and emphasize informative updates throughout the alignment process. Extensive experiments on various baseline models, including image generation models on SD1.5, SDXL, and video generation models CogVideoX-2B, CogVideoX-5B, and Wan2.1-1.3B, demonstrate that our SIPO consistently promotes stabilized training and outperforms existing alignment methods, with meticulous adjustments on parameters.\nOverall, these results highlight the importance of timestep-aware alignment and and provide valuable guidelines for improved preference optimization in diffusion models."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Diffusion",
"DPO",
"image generate",
"video generate"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/38259828575fb003f42adcd3a998c716c9e1533f.pdf"
},
"primary_area": {
"value": "generative models"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/0e37b8ed7b6e3832740d3b2b86fda314d0c201d4.zip"
},
"title": {
"value": "SIPO: Stabilized and Improved Preference Optimization for Aligning Diffusion Models"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "06upBSlAUy",
"id": "06upBSlAUy",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission23238/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759896824854,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission23238/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission23238/Authors"
]
} | |
2,026 | 072P11r1wu | [
0,
10,
2,
2
] | [
{
"content": "Benign and harmful overfitting have been extensively studied in the past few years in many settings and models. More recently, there has been interest in analyzing benign overfitting in simple transformers. This work aims to extend the previous works on benign/harmful overfitting in transformers b... | {
"cdate": 1758292502092,
"content": {
"TLDR": {
"value": "We present generalization bounds for a two-layer Transformer under benign overfitting and harmful overfitting."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025understanding,\ntitle={Understanding Generalization in Transformers: Error Bounds and Training Dynamics Under Benign and Harmful Overfitting},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=072P11r1wu},\nnote={under review}\n}"
},
"abstract": {
"value": "Transformers serve as the foundational architecture for many successful large-scale models, demonstrating the ability to overfit the training data while maintaining strong generalization on unseen data, a phenomenon known as benign overfitting. However, existing research has not sufficiently explored generalization and training dynamics of transformers under benign overfitting. This paper addresses this gap by analyzing a two-layer transformer's training dynamics, convergence, and generalization under labeled noise. Specifically, we present generalization error bounds for benign and harmful overfitting under varying signal-to-noise ratios (SNR), where the training dynamics are categorized into three distinct stages, each with its corresponding error bounds. Additionally, we conduct extensive experiments to identify key factors in transformers that influence test losses. Our experimental results align closely with the theoretical predictions, validating our findings."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Transformer",
"Benign overfiting",
"Feature learning theory",
"Generalization error bounds",
"Signal-to-noise ratio"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/231fabf71a2994eee63b8ef73378fafe7e9c94c6.pdf"
},
"primary_area": {
"value": "learning theory"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/005086072521ed959ef9357196c2aea60a6faaab.zip"
},
"title": {
"value": "Understanding Generalization in Transformers: Error Bounds and Training Dynamics Under Benign and Harmful Overfitting"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "072P11r1wu",
"id": "072P11r1wu",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission18976/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897069595,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission18976/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission18976/Authors"
]
} | |
2,026 | 073WQjmWKU | [
6,
6,
4,
6,
8,
4
] | [
{
"content": "This paper presents COMPACT, a data-efficient visual instruction tuning (VIT) framework that synthesizes training samples with controlled compositional complexity. The authors introduce the k-value, representing the number of atomic visual capabilities (e.g., object recognition, spatial reasoning)... | {
"cdate": 1757812256580,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025compact,\ntitle={{COMPACT}: {COMP}ositional Atomic-to-Complex Visual Capability Tuning},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=073WQjmWKU},\nnote={under review}\n}"
},
"abstract": {
"value": "Visual instruction tuning (VIT) datasets consist of randomly sampled image-question pairs without regard to the informativeness of each pair. Recent dataset selection methods have shown that a small fraction of such datasets enriched with informative samples can lead to efficient finetuning of Multimodal Large Language Models. In this work, we explore the impact of task complexity on informative data curation and introduce COMPACT (COMPositional Atomic-to-complex Visual Capability Tuning), a VIT data recipe that scales training sample complexity by combining multiple atomic visual capabilities in a single training example. Concretely, we synthesize rich and informative text questions for each image, allowing us to significantly reduce the number of training examples required for effective visual instruction tuning. COMPACT demonstrates superior data efficiency compared to existing data reduction methods. When applied to the LLaVA-665K VIT dataset, COMPACT reduces the data budget by 90% while still achieving 100.2% of the full VIT performance (compared to only 97.5% by the state-of-the-art method) across eight multimodal benchmarks. Further, training on the same COMPACT data even improves performance compared to training with full-scale data on particularly complex benchmarks such as MM-Vet (+8.6%) and MMStar (+2.9%). COMPACT offers a scalable and efficient synthetic data generation recipe to improve on visual language tasks."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Complexity",
"Compositionality",
"Visual instruction tuning"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/12d65dd15508e5a6a60d5da82927f4e8a13e8f75.pdf"
},
"primary_area": {
"value": "foundation or frontier models, including LLMs"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "COMPACT: COMPositional Atomic-to-Complex Visual Capability Tuning"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "073WQjmWKU",
"id": "073WQjmWKU",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission4933/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759898004339,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission4933/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission4933/Authors"
]
} | |
2,026 | 075TvkpZEK | [
2,
4,
8,
2
] | [
{
"content": "This paper proposes an optimization algorithm SMARAN for deep learning. SMARAN has two main characteristics, the first is that it normalizes the gradient before updating the first-order momentum, and the second is that it adopts the objective function value to update the second-order momentum. The... | {
"cdate": 1758257347389,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025smaran,\ntitle={{SMARAN}: Closing the Generalization Gap with Performance Driven Optimization Method},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=075TvkpZEK},\nnote={under review}\n}"
},
"abstract": {
"value": "Optimization methods have evolved significantly by introducing various learning rate scheduling techniques and adaptive learning strategies. Although these methods have achieved faster convergence, they often struggle to generalize well to unseen data compared to traditional approaches such as Stochastic Gradient Descent (SGD) with momentum. Adaptive methods such as Adam store each parameter's first and second moments of gradients, which can be memory-intensive. To address these challenges, we propose a novel SMARAN optimization method that adjusts the learning rate based on the model's performance rather than the objective function's curvature. This approach is particularly effective for minimizing stochastic loss functions, standard in deep learning models. Traditional gradient-based methods may get stuck in regions where the gradient vanishes, such as plateaus or local minima. Therefore, instead of only depending on the gradient, we use the model's performance to estimate the appropriate step size. We performed extensive experiments on standard vision benchmarks, and the generalization trends observed with SMARAN demonstrate compelling distinctions relative to adaptive and non-adaptive optimizers."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Optimization",
"Gradient decent",
"Learning rate scheduler",
"Regularization"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/51bea55276c3d93f4d63af8d07abe5cf281cc86d.pdf"
},
"primary_area": {
"value": "optimization"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "SMARAN: Closing the Generalization Gap with Performance Driven Optimization Method"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "075TvkpZEK",
"id": "075TvkpZEK",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission15935/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897272046,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission15935/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission15935/Authors"
]
} | |
2,026 | 07R3pHnBqc | [
2,
0,
4,
2
] | [
{
"content": "This paper proposes Instruction Agent, a training-free GUI automation framework that uses a single expert demonstration, aiming to execute long-horizon and complex tasks. The instructor module ensures the agent follows the instruction, while the verifier and backtracker aim to ensure robustness. T... | {
"cdate": 1757981401623,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@misc{\nli2025instruction,\ntitle={Instruction Agent: Enhancing Agent with Expert Demonstration},\nauthor={Yinheng Li and Hailey Hultquist and Justin Wagle and Kazuhito Koishida},\nyear={2025},\nurl={https://openreview.net/forum?id=07R3pHnBqc}\n}"
},
"abstract": {
"value": "Graphical user interface (GUI) agents have advanced rapidly but still struggle with complex tasks involving novel UI elements, long-horizon actions, and personalized trajectories. In this work, we introduce Instruction Agent, a GUI agent that leverages expert demonstrations to solve such tasks, enabling completion of otherwise difficult workflows. Given a single demonstration, the agent extracts step-by-step instructions and executes them by strictly following the trajectory intended by the user, which avoids making mistakes during execution. The agent leverages the verifier and backtracker modules further to improve robustness. Both modules are critical to understand the current outcome from each action and handle unexpected interruptions(such as pop-up windows) during execution. Our experiments show that Instruction Agent achieves a 60% success rate on a set of tasks in OSWorld that all top-ranked agents failed to complete. The Instruction Agent offers a practical and extensible framework, bridging the gap between current GUI agents and reliable real-world GUI task automation."
},
"anonymous_url": null,
"authorids": {
"value": [
"~Yinheng_Li2",
"~Hailey_Hultquist1",
"~Justin_Wagle1",
"~Kazuhito_Koishida1"
]
},
"authors": {
"value": [
"Yinheng Li",
"Hailey Hultquist",
"Justin Wagle",
"Kazuhito Koishida"
]
},
"code_of_ethics": null,
"keywords": {
"value": [
"GUI Agents",
"Expert Demonstrations",
"Human in the Loop",
"Test-Time Automation",
"Backtracking"
]
},
"no_acknowledgement_section": null,
"paperhash": {
"value": "li|instruction_agent_enhancing_agent_with_expert_demonstration"
},
"pdf": {
"value": "/pdf/fdf65d865fac30aa8de77cb5dab6e301148a2621.pdf"
},
"primary_area": {
"value": "applications to robotics, autonomy, planning"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Instruction Agent: Enhancing Agent with Expert Demonstration"
},
"venue": {
"value": "ICLR 2026 Conference Withdrawn Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Withdrawn_Submission"
}
},
"forum": "07R3pHnBqc",
"id": "07R3pHnBqc",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission6408/-/Full_Submission",
"ICLR.cc/2026/Conference/-/Withdrawn_Submission"
],
"license": "CC BY 4.0",
"mdate": 1762926589336,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission6408/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission6408/Authors"
]
} | |
2,026 | 07S1CPoQYP | [
6,
2,
2,
2
] | [
{
"content": "This paper investigates how fMRI recordings can be used to fine-tune large language models (LLMs) toward human brain activity. The authors propose a dual-objective framework combining standard language modeling with brain alignment, leveraging over 50 hours of naturalistic movie-watching fMRI data... | {
"cdate": 1758149730341,
"content": {
"TLDR": {
"value": "We show that brain-informed training of language models, using dual objectives and scaling across data, models, and subjects, yields robust and generalizable alignment with human brain activity beyond baselines."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025braininformed,\ntitle={Brain-Informed Language Model Training Enables Scalable and Generalizable Alignment with Human Brain Activity},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=07S1CPoQYP},\nnote={under review}\n}"
},
"abstract": {
"value": "Language models (LMs) provide rich representational spaces that partially align with neural activity during naturalistic experiences such as movie watching. Yet leveraging brain recordings to actively guide LM training remains underexplored. Here, we address this question by investigating whether functional MRI recordings can guide LLM training by aligning language representations with brain dynamics. Using over 50 hours of fMRI data from six participants watching Friends, plus 10 hours of held-out movies, we augmented pre-trained and randomly initialized LMs with a brain alignment module and compared multiple training strategies. Our results show three main findings. First, brain-informed fine-tuning consistently outperforms text-only baselines and brain-from-scratch models, with voxel-level gains that scale with both model size (GPT-2 124M, LLaMA-2 7B) and training duration (1–40 hours). These improvements generalize across participants and out-of-sample movies, yielding robust cross-subject and cross-stimulus encoding. Second, a dual-objective loss that balances language modeling with brain alignment surpasses brain-only optimization, producing more stable and generalizable encoders. Finally, brain supervision enriches LM representations with multisensory inductive biases: brain-fine-tuned models outperform unimodal baselines on VL-Commonsense, better capturing perceptual and associative properties (e.g., color, shape, co-occurrence) that text-only training underrepresents. Together, these results establish cortical dynamics as an effective supervisory signal, enabling scalable, generalizable, and brain-aligned LMs that internalize aspects of human-like multimodal representation."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Language models",
"fMRI",
"Encoding models",
"naturalistic stimulus",
"Representation Learning",
"Multimodal Learning",
"Low-Rank Adaptation (LoRA)",
"Transfer Learning",
"Neuroscience-Informed AI",
"Scalability",
"Generalizability"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/88e12edfd2a5e85bc41b8134b798939c093b5c2d.pdf"
},
"primary_area": {
"value": "applications to neuroscience & cognitive science"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Brain-Informed Language Model Training Enables Scalable and Generalizable Alignment with Human Brain Activity"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "07S1CPoQYP",
"id": "07S1CPoQYP",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission9930/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897684766,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission9930/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission9930/Authors"
]
} | |
2,026 | 07o2iouN1Y | [
2,
4,
6,
2
] | [
{
"content": "The paper solves Nash equilibrium in two-player zero-sum extensive-form games by adding additional regularization. By switching the reference strategy periodically, the algorithm converges to the NE of the original game rather than the regularized game. The paper proves convergence theoretically a... | {
"cdate": 1757304899303,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025nash,\ntitle={Nash Policy Gradient: A Policy Gradient Method with Iteratively Refined Regularization for Finding Nash Equilibria},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=07o2iouN1Y},\nnote={under review}\n}"
},
"abstract": {
"value": "Finding Nash equilibria in imperfect-information games remains a central challenge in multi-agent reinforcement learning. While regularization-based methods have recently achieved last-iteration convergence to a regularized equilibrium, they require the regularization strength to shrink toward zero to approximate a Nash equilibrium, often leading to unstable learning in practice. Instead, we fix the regularization strength at a large value for robustness and achieve convergence by iteratively refining the reference policy. Our main theoretical result shows that this procedure guarantees strictly monotonic improvement and convergence to an exact Nash equilibrium in two-player zero-sum games, without requiring a uniqueness assumption. Building on this framework, we develop a practical algorithm, *Nash Policy Gradient* (NashPG), which preserves the generalizability of policy gradient methods while relying solely on the current and reference policies. Empirically, NashPG achieves comparable or lower exploitability than prior model-free methods on classic benchmark games and scales to large domains such as *Battleship* and *No-Limit Texas Hold'em*, where NashPG consistently attains higher Elo ratings."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Multi-agent reinforcement learning",
"policy gradient",
"game theory",
"Nash equilibria"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/8c6be1ccaa4cfd6f0abd440e6f11f0b609242681.pdf"
},
"primary_area": {
"value": "reinforcement learning"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Nash Policy Gradient: A Policy Gradient Method with Iteratively Refined Regularization for Finding Nash Equilibria"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "07o2iouN1Y",
"id": "07o2iouN1Y",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission2942/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759898118185,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission2942/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission2942/Authors"
]
} | |
2,026 | 084SvT55yk | [
10,
4,
6
] | [
{
"content": "Existing neural CO solvers either ensure local feasibility but lack global awareness (LC) or produce global predictions with constraint violations (GP). Current adaptive expansion is only an external wrapper with limited effectiveness.\nNEXCO makes adaptive expansion native through CO-specific mas... | {
"cdate": 1756822893695,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025native,\ntitle={Native Adaptive Solution Expansion for Diffusion-based Combinatorial Optimization},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=084SvT55yk},\nnote={under review}\n}"
},
"abstract": {
"value": "One central challenge in Neural Combinatorial Optimization (NCO) is handling hard constraints efficiently. Beyond the two classic paradigms, i.e., Local Construction (LC), which sequentially builds feasible solutions but scales poorly, and Global Prediction (GP), which produces one-shot heatmaps yet struggles with constraint conflicts, the recently proposed Adaptive Expansion (AE) shares the advantages of both by progressively growing partial solutions with instance-wise global awareness.\nHowever, existing realizations bolt AE onto external GP predictors, so their solution quality is bounded by the backbone and their inference cost scales with repeated global calls.\nIn this paper, we fundamentally rethink adaptive expansion and make it native to a generative model, acting as its intrinsic decoding principle rather than an external wrapper.\nWe propose NEXCO, a CO-specific masked diffusion framework that turns adaptive expansion into the model’s own iterative unmasking process.\nSpecifically, it involves a solution-expansion training procedure with a time-agnostic GNN denoiser, which learns diffusion trajectories between fully masked solutions and ground-truth solutions.\nWith the trained time-agnostic denoiser, we introduce a novel solution expansion scheme at the solving stage, enabling adaptive control over the intermediate solution states. \nIt is achieved by constructing candidate sets according to confidence scores and applying feasibility projection to expand the solution while respecting constraints. \nIn this way, ``adaptive\" is not an afterthought but the decoding itself: intermediate diffusion states are meaningful partial solutions and progress is instance-adaptive rather than schedule-bound.\nExtensive experiments on representative CO problems show that NEXCO achieves approximately 50\\% improvement in solution quality and up to $4\\times$ faster inference compared to prior state-of-the-art solvers."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"mask diffusion model",
"neural combinatorial optimization"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/ba6d77226d561157026830477edbe78117c8becb.pdf"
},
"primary_area": {
"value": "learning on graphs and other geometries & topologies"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Native Adaptive Solution Expansion for Diffusion-based Combinatorial Optimization"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "084SvT55yk",
"id": "084SvT55yk",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission903/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759898236270,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission903/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission903/Authors"
]
} | |
2,026 | 08EyZzhgl1 | [
2,
4,
2,
4
] | [
{
"content": "This paper presents TextME, a text‑only training framework for modality expansion that eliminates the need for paired multimodal data by leveraging the “consistent modality gap” property of pretrained encoders. TextME first pre‑computes a constant offset between text and non‑text embeddings for ea... | {
"cdate": 1758266488620,
"content": {
"TLDR": {
"value": "TextMEunifies specialized modalities without paired supervision by training text-only projectors and applying centering offsets to bridge the modality gap at inference."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025textme,\ntitle={Text{ME}: Text-only Training for Modality Expansion via {LLM} Space Pivoting},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=08EyZzhgl1},\nnote={under review}\n}"
},
"abstract": {
"value": "Expanding multimodal representations to novel modalities is constrained by reliance on large-scale paired datasets (e.g., text–image, text–audio, text–3D, text–molecule), which are costly and often infeasible in domains requiring expert annotation such as medical imaging, 3D modeling, and molecular analysis. We introduce TextME, the first framework for text-only modality expansion that removes paired data requirements. Our method leverages the universal geometric properties of pre-trained encoders—consistent modality gaps—which enable zero-shot cross-modal transfer once embedding spaces satisfy these properties. We empirically verify that these hold across audio, 3D, X-ray, and molecular domains, enabling effective cross-modal tasks without paired supervision. Furthermore, we evaluated LLM and multimodal text encoders to determine which is more effective as a unified anchor space. Experiments show that TextME achieves 88.2% of paired-data performance in zero-shot classification and cross-modal retrieval, while also supporting emergent capabilities between unseen modality pairs (e.g., audio-to-3D, molecule-to-image). These results highlight text-only modality expansion as a practical and scalable path toward foundation models spanning arbitrary modalities."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"multimodal learning",
"modality expansion",
"text-only training",
"modality gap",
"cross-modal retrieval",
"representation alignment"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/d88a5b6ee725f802837c6cc50d23eae58e76be85.pdf"
},
"primary_area": {
"value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "TextME: Text-only Training for Modality Expansion via LLM Space Pivoting"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "08EyZzhgl1",
"id": "08EyZzhgl1",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission16593/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897230825,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission16593/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission16593/Authors"
]
} | |
2,026 | 08FTG45E9m | [
4,
6,
2,
2
] | [
{
"content": "The paper introduces Hermes, a multi-scale spatial-temporal hypergraph network for stock time series forecasting. The model aims to jointly model inter-industry lead-lag structures and multi-scale temporal dependencies. It incorporates a hyperedge-based moving aggregation module and a multi-scale ... | {
"cdate": 1756734867685,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@misc{\nqiu2025multiscale,\ntitle={Multi-Scale Spatial-Temporal Hypergraph Network with Lead-Lag Structures for Stock Time Series Forecasting},\nauthor={Xiangfei Qiu and Liu Yang and Hanyin Cheng and Xingjian Wu and Rongjia Wu and Zhang Zhigang and Tu ding and Chenjuan Guo and Bin Yang and Christian S. Jensen and Jilin Hu},\nyear={2025},\nurl={https://openreview.net/forum?id=08FTG45E9m}\n}"
},
"abstract": {
"value": "Time series forecasting occurs in a range of financial applications providing essential decision-making support to investors, regulatory institutions, and analysts. Unlike multivariate time series from other domains, stock time series exhibit industry correlation. Exploiting this kind of correlation can improve forecasting accuracy. However, existing methods based on hypergraphs can only capture industry correlation relatively superficially. These methods face two key limitations: they do not fully consider inter-industry lead-lag interactions, and they do not model multi-scale information within and among industries. This study proposes the Hermes framework for stock time series forecasting that aims to improve the exploitation of industry correlation by eliminating these limitations. The framework integrates moving aggregation and multi-scale fusion modules in a hypergraph network. Specifically, to more flexibly capture the lead-lag relationships among industries, Hermes proposes a hyperedge-based moving aggregation module. This module incorporates a sliding window and utilizes dynamic temporal aggregation operations to consider lead-lag dependencies among industries. Additionally, to effectively model multi-scale information, Hermes employs cross-scale, edge-to-edge message passing to integrate information from different scales while maintaining the consistency of each scale. Experimental results on multiple real-world stock datasets show that Hermes outperforms existing state-of-the-art methods in both efficiency and accuracy."
},
"anonymous_url": null,
"authorids": {
"value": [
"~Xiangfei_Qiu1",
"~Liu_Yang22",
"~Hanyin_Cheng1",
"~Xingjian_Wu1",
"~Rongjia_Wu2",
"~Zhang_Zhigang1",
"~Tu_ding1",
"~Chenjuan_Guo1",
"~Bin_Yang4",
"~Christian_S._Jensen1",
"~Jilin_Hu1"
]
},
"authors": {
"value": [
"Xiangfei Qiu",
"Liu Yang",
"Hanyin Cheng",
"Xingjian Wu",
"Rongjia Wu",
"Zhang Zhigang",
"Tu ding",
"Chenjuan Guo",
"Bin Yang",
"Christian S. Jensen",
"Jilin Hu"
]
},
"code_of_ethics": null,
"keywords": {
"value": [
"time series",
"time series forecasting"
]
},
"no_acknowledgement_section": null,
"paperhash": {
"value": "qiu|multiscale_spatialtemporal_hypergraph_network_with_leadlag_structures_for_stock_time_series_forecasting"
},
"pdf": {
"value": "/pdf/bc21800b8aa59aafd1af1d5aa573e3f8aab3d55b.pdf"
},
"primary_area": {
"value": "learning on time series and dynamical systems"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Multi-Scale Spatial-Temporal Hypergraph Network with Lead-Lag Structures for Stock Time Series Forecasting"
},
"venue": {
"value": "ICLR 2026 Conference Withdrawn Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Withdrawn_Submission"
}
},
"forum": "08FTG45E9m",
"id": "08FTG45E9m",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission308/-/Full_Submission",
"ICLR.cc/2026/Conference/-/Withdrawn_Submission"
],
"license": "CC BY 4.0",
"mdate": 1763294467721,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission308/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission308/Authors"
]
} | |
2,026 | 08KOxSjRyj | [
4,
2,
4,
2
] | [
{
"content": "The paper introduces LongEmotion, a long-context benchmark for evaluating LLMs Emotional Intelligence (EI) across six task: Emotion Classification, Emotion Detection, Emotion QA, Emotion Conversation, Emotion Summary, and Emotion Expression. Moreover， this paper propose RAG and CoEM frameworks to ... | {
"cdate": 1758170614961,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025longemotion,\ntitle={LongEmotion: Measuring Emotional Intelligence of Large Language Models in Long-Context Interaction},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=08KOxSjRyj},\nnote={under review}\n}"
},
"abstract": {
"value": "Large language models (LLMs) make significant progress in Emotional Intelligence (EI) and long-context understanding. However, existing benchmarks tend to overlook certain aspects of EI in long-context scenarios, especially under $\\textit{realistic, practical settings}$ where interactions are lengthy, diverse, and often noisy. To move towards such realistic settings, we present $\\textit{LongEmotion}$, a benchmark specifically designed for long-context EI tasks. It covers a diverse set of tasks, including $\\textbf{Emotion Classification}$, $\\textbf{Emotion Detection}$, $\\textbf{Emotion QA}$, $\\textbf{Emotion Conversation}$, $\\textbf{Emotion Summary}$, and $\\textbf{Emotion Expression}$. On average, the input length for these tasks reaches 8${,}$777 tokens, with long-form generation required for $\\textit{Emotion Expression}$. To enhance performance under realistic constraints, we incorporate Retrieval-Augmented Generation ($\\textit{RAG}$) and Collaborative Emotional Modeling ($\\textit{CoEM}$), and compare them with standard prompt-based methods. Unlike conventional approaches, our $\\textit{RAG}$ method leverages both the conversation context and the large language model itself as retrieval sources, avoiding reliance on external knowledge bases. The $\\textit{CoEM}$ method further improves performance by decomposing the task into five stages, integrating both retrieval augmentation and limited knowledge injection. Experimental results show that both $\\textit{RAG}$ and $\\textit{CoEM}$ consistently enhance EI-related performance across most long-context tasks, advancing LLMs toward more $\\textit{practical and real-world EI applications}$. Furthermore, we conduct a detailed case study on the performance comparison among GPT series models, the application of CoEM in each stage and its impact on task scores, and the advantages of the LongEmotion dataset in advancing EI. All of our code and datasets will be open-sourced, which can be viewed at the anonymous repository link https://anonymous.4open.science/r/anonymous-578B."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Emotional Intelligence",
"Long-Context"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/5c3e0c117d155281a7979617728393553fab82e5.pdf"
},
"primary_area": {
"value": "datasets and benchmarks"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/bcac58c19d6cdce0babe4a787d2efe6dff891815.zip"
},
"title": {
"value": "LongEmotion: Measuring Emotional Intelligence of Large Language Models in Long-Context Interaction"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "08KOxSjRyj",
"id": "08KOxSjRyj",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission10412/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897652232,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission10412/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission10412/Authors"
]
} | |
2,026 | 08pxmTLKTT | [
2,
4,
4,
6
] | [
{
"content": "The paper proposes to better address the problem of object ambiguity in interactive segmentation (IS) models with SmartSAM method. To achieve it, an agent generates a few branches with interactions (positive/negative click or bbox), considering the first user interaction, to produce candidate mask... | {
"cdate": 1758013764632,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025smartsam,\ntitle={Smart{SAM}: Segment Ambiguous Objects like Smart Annotators},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=08pxmTLKTT},\nnote={under review}\n}"
},
"abstract": {
"value": "Segment Anything Model (SAM) often encounters ambiguity in interactive segmentation, where insufficient user interaction leads to inaccurate segmentation of the target object. Existing approaches primarily address ambiguity through repeated human-model interactions, which are time-consuming due to the inherent latency of human responses. To reduce human efforts, we propose a novel interactive segmentation framework that leverages the model’s inherent capabilities to effectively segment ambiguous objects.\nOur key idea is to create an annotator-like agent to interact with the model. The resulting SmartSAM method mimics intelligent human annotators, resolving ambiguity with a single click and one reference instance. The agent generates multiple prompts around the initial click to simulate diverse annotator behaviors and refines the output masks by iteratively adding click chains in uncertain regions, thereby producing a set of candidate masks. Finally, the agent selects the mask that most closely aligns with the user’s intent, as indicated by the reference instance. Furthermore, we formalize the agent’s behavior as a fuzzy regression problem by quantifying ambiguity using fuzzy entropy. We demonstrate that our agent yields lower entropy than traditional methods, and we establish robustness and sufficiency theorems to ensure effective, human-like decision-making within a bounded range of actions. We evaluate our approach on multiple segmentation benchmarks and demonstrate its superiority over state-of-the-art methods."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Ambiguity",
"Segment Anything Model",
"Interactive Segmentation"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/820ea9ba35674f9e0d8e8f0e96b98688d2ecfb4e.pdf"
},
"primary_area": {
"value": "applications to computer vision, audio, language, and other modalities"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "SmartSAM: Segment Ambiguous Objects like Smart Annotators"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "08pxmTLKTT",
"id": "08pxmTLKTT",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission7273/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897862687,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission7273/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission7273/Authors"
]
} | |
2,026 | 08tuDzMDEn | [
4,
4,
6,
4
] | [
{
"content": "This paper studies the task of counterargument generation and introduces a persona-based approach with Tree-of-Thought (ToT) content planning. Specifically, given an original post (OP), the system first constructs three distinct personas, each representing a unique perspective. These personas then... | {
"cdate": 1758269143514,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025ptcg,\ntitle={{PTCG}: Persona-guided Tree-based Counterargument Generation},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=08tuDzMDEn},\nnote={under review}\n}"
},
"abstract": {
"value": "The ability to generate counterarguments is important for fostering critical thinking, balanced discourse, and informed decision-making.\nHowever, existing approaches typically produce only a single counterargument, thereby overlooking the diversity and persuasiveness required in real-world debates.\nThis limitation is critical, as the same topic may persuade different individuals only when framed from distinct perspectives.\nTo address this limitation, we propose Persona-guided Tree-based Counterargument Generation (PTCG), a framework that combines Tree-of-Thoughts–inspired step-wise generation and pruning with speaker persona selection.\nBy estimating the author’s persona from the original argument and incorporating speaker personas representing distinct perspectives, the framework operationalizes perspective-taking, enabling reasoning from multiple standpoints and supporting the generation of diverse counterarguments.\nWe propose a tree-based procedure that generates plans, selects the best, and produces multiple speaker persona-specific counterarguments, from which the most effective are chosen.\nWe evaluate PTCG through a comprehensive multi-faceted setup, combining Large Language Model (LLM)-as-a-Judge, classifier-based assessment, and human evaluations.\nOur experimental results show that PTCG substantially improves both the diversity and persuasiveness of counterarguments compared to baselines.\nThese findings highlight the effectiveness of adaptive persona integration in boosting diversity and strengthening persuasiveness."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Counterargument",
"Generation",
"Persona"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/91dbff0a08990793b704ae81302b0af1ddc72c2c.pdf"
},
"primary_area": {
"value": "applications to computer vision, audio, language, and other modalities"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/3269496ebac5f0c8275f3b4891cbd3a2309344fb.zip"
},
"title": {
"value": "PTCG: Persona-guided Tree-based Counterargument Generation"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "08tuDzMDEn",
"id": "08tuDzMDEn",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission16826/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897216977,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission16826/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission16826/Authors"
]
} | |
2,026 | 09FE8nv4sV | [
6,
4,
4,
4
] | [
{
"content": "This paper introduces \"MILP-Retrieval,\" a novel framework to address the critical problem of data scarcity for training data-driven Mixed-Integer Linear Programming (MILP) solvers. The authors argue that existing generative methods (e.g., VAEs, diffusion models) are highly inefficient, as they r... | {
"cdate": 1758211996865,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025targeted,\ntitle={Targeted {MILP} Instance Generation via Formulation Code Retrieval},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=09FE8nv4sV},\nnote={under review}\n}"
},
"abstract": {
"value": "Efficient and controllable data generation is critical for improving the performance of data-driven Mixed-Integer Linear Programming (MILP) solvers, especially in applications facing data scarcity. However, existing MILP instance generation methods typically require training a separate model for each problem class, which can be computationally intensive and does not allow for the generation of instances with varying sizes and solution difficulties. To address these challenges, we introduce MILP-Retrieval, a framework for targeted MILP instance generation via formulation code retrieval. We first build a diverse MILP library that includes multiple modalities and use it to pretrain an MILP embedding model. Based on the output of this embedding model, we propose a novel similarity metric that accurately measures the similarity between instances of different sizes within the same problem class. MILP-Retrieval leverages this new metric to retrieve the formulation code of a target instance and further tune it. Experimental results demonstrate the effectiveness of generating MILP instances through formulation code retrieval, with the ability to control both the scale and difficulty of the generated instances. This approach provides a novel perspective on MILP instance generation and opens up new possibilities for learning-based solvers."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Mixed-Integer Linear Programming",
"Combinatorial Optimization",
"MILP Instance Generation"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/68b05fde92bc24ee1cdc82d5c27e9096a8eaed7c.pdf"
},
"primary_area": {
"value": "optimization"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Targeted MILP Instance Generation via Formulation Code Retrieval"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "09FE8nv4sV",
"id": "09FE8nv4sV",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission12955/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897474393,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission12955/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission12955/Authors"
]
} | |
2,026 | 09Nj40ScvC | [
2,
6,
4
] | [
{
"content": "The paper proposes a heuristic to select preference pairs to train PRM. Meanwhile, it also modifies the advantage function of original GRPO to adapt to process reward settings.",
"id": "rgdCO8yLHy",
"rating": 2
},
{
"content": "The paper presents a novel reinforcement learning fram... | {
"cdate": 1758348612106,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025preferencebased,\ntitle={Preference-Based Process Reward Model for Robust Mathematical Reasoning},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=09Nj40ScvC},\nnote={under review}\n}"
},
"abstract": {
"value": "Process reward models (PRMs) have emerged as a promising approach to guide LLMs by providing step-wise supervision, but traditional methods often rely on heuristic search strategies like Monte Carlo Tree Search (MCTS), which introduce bias and limit generalization. In this work, we propose a reinforcement learning framework guided by a Preference-Based Process Reward Model (PPRM) , which provides step-wise supervision to refine reasoning trajectories. We first employ MCTS to estimate and select chosen and rejected rollouts, thereby constructing a high-quality step-level dataset. Our PPRM is trained on Bradley-Terry loss function, which mitigates the bias introduced by the heuristic search strategies of MCTS by leveraging preference-based learning. To enable effective RL training with PPRM, we enhance Group Relative Policy Optimization (GRPO) by introducing a robust advantage estimator that better captures the structure of preference-based process reward model enabling stable and efficient policy optimization. Experimental results on ProcessBench and best-of-n strategy demonstrate that our approach achieves $2$-$3\\%$ improvement in intermediate step accuracy compared to existing methods for complex reasoning processes, thereby improving the reasoning accuracy of the policy model across several key reasoning benchmarks."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Process Reward Model",
"Reinforcement Learning",
"Monte Carlo Tree Search"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/1a96cc56ba0b8dae8c095aa61356d3af1319282c.pdf"
},
"primary_area": {
"value": "foundation or frontier models, including LLMs"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Preference-Based Process Reward Model for Robust Mathematical Reasoning"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "09Nj40ScvC",
"id": "09Nj40ScvC",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission23801/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759896796319,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission23801/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission23801/Authors"
]
} | |
2,026 | 09YSBymX6O | [
6,
4,
8
] | [
{
"content": "The authors propose using spatial point processes as a self-supervision prior that explicitly models spatial distributions of objects to address the gap in previous un- and self-supervised methods that miss the spatial correlations. Thus, the paper proposes spatially informed variational autoenco... | {
"cdate": 1758200193649,
"content": {
"TLDR": {
"value": "We present spatially informed variational autoencoders that use stochastic point processes to learn interpretable spatial patterns from images."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025spatially,\ntitle={Spatially Informed Autoencoders for Interpretable Visual Representation Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=09YSBymX6O},\nnote={under review}\n}"
},
"abstract": {
"value": "We introduce spatially informed variational autoencoders (SI-VAE) as self-supervised deep-learning models that use stochastic point processes to predict spatial organization patterns from images. Existing approaches to learning visual representations based on variational autoencoders (VAE) struggle to capture spatial correlations between objects or events, focusing instead on pixel intensities. We address this limitation by incorporating a point-process likelihood, derived from the Papangelou conditional intensity, as a self-supervision target. This results in a hybrid model that learns statistically interpretable representations of spatial localization patterns and enables zero-shot conditional simulation directly from images. Experiments with synthetic images show that SI-VAE improve the classification accuracy of attractive, repulsive, and uncorrelated point patterns from 48% (VAE) to over 80% in the worst case and 90% in the best case, while generalizing to unseen data. We apply SI-VAE to a real-world microscopy data set, demonstrating its use for studying the spatial organization of proteins in human cells and for using the representations in downstream statistical analysis."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"autoencoder",
"visual representation",
"point process",
"conditional simulation",
"interpretable machine learning",
"self supervision",
"spatial statistics"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/7748035b13c9cd12469e6ea6b4f2b5d09a0bd09d.pdf"
},
"primary_area": {
"value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Spatially Informed Autoencoders for Interpretable Visual Representation Learning"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "09YSBymX6O",
"id": "09YSBymX6O",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission11483/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897572558,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission11483/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission11483/Authors"
]
} | |
2,026 | 09lmwhDqZ3 | [
6,
4,
6,
6
] | [
{
"content": "This paper focuses on the task of automatic formalization in theorem proving, which currently faces two major challenges: model hallucination and the semantic gap caused by ambiguous or missing premises in natural language descriptions. To address these issues, the authors propose a framework call... | {
"cdate": 1758191085758,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025automated,\ntitle={Automated Formalization via Conceptual Retrieval-Augmented {LLM}s},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=09lmwhDqZ3},\nnote={under review}\n}"
},
"abstract": {
"value": "Interactive theorem provers (ITPs) require manual formalization, which is labor-intensive and demands expert knowledge. While automated formalization offers a potential solution, it faces two major challenges: model hallucination (e.g., undefined predicates, symbol misuse, and version incompatibility) and the semantic gap caused by ambiguous or missing premises in natural language descriptions. To address these issues, we propose CRAMF, a Concept-driven Retrieval-Augmented Mathematical Formalization framework. CRAMF enhances LLM-based autoformalization by retrieving formal definitions of core mathematical concepts, providing contextual grounding during code generation. However, applying retrieval-augmented generation (RAG) in this setting is non-trivial due to the lack of structured knowledge bases, the polymorphic nature of mathematical concepts, and the high precision required in formal retrieval. We introduce a framework for automatically constructing a concept-definition knowledge base from Mathlib4, the standard mathematical library for the Lean 4 theorem prover, indexing over 26,000 formal definitions and 1,000+ core mathematical concepts. To address conceptual polymorphism, we propose contextual query augmentation with domain- and application-level signals. In addition, we design a dual-channel hybrid retrieval strategy with reranking to ensure accurate and relevant definition retrieval. Experiments on miniF2F, ProofNet, and our newly proposed AdvancedMath benchmark show that CRAMF can be seamlessly integrated into LLM-based autoformalizers, yielding consistent improvements in translation accuracy—achieving up to 62.1% and an average of 29.9% relative improvement."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Autoformalization",
"Retrieval-augmented Generation"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/4089192f22e6b5c7e3ec876501c9474f4e95d04d.pdf"
},
"primary_area": {
"value": "foundation or frontier models, including LLMs"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/a8c5ee057e703486a6d99918ad99eceeec8faff6.zip"
},
"title": {
"value": "Automated Formalization via Conceptual Retrieval-Augmented LLMs"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "09lmwhDqZ3",
"id": "09lmwhDqZ3",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission11148/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897604259,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission11148/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission11148/Authors"
]
} | |
2,026 | 0A2rXt5SAy | [
2,
6,
4,
4
] | [
{
"content": "This paper proposes a method that solves the pipeline scheduling problem using mixed-integer linear programming (MILP), treating activation offloading as a decision variable. It models whether activations are offloaded or retained in GPU memory and enforces constraints on data dependencies, resour... | {
"cdate": 1758208530100,
"content": {
"TLDR": {
"value": "Use Mathematical Programming to model Pipeline Parallelism with Offloading to balance efficiency and memory requirement."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025optpipe,\ntitle={OptPipe: Memory- and Scheduling-Optimized Pipeline Parallelism for {LLM} Training},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0A2rXt5SAy},\nnote={under review}\n}"
},
"abstract": {
"value": "Pipeline parallelism (PP) has become a standard technique for scaling large language model (LLM) training across multiple devices. However, despite recent progress in reducing memory consumption through activation offloading, existing approaches remain largely heuristic and coarse-grained, often overlooking the fine-grained trade-offs between memory, computation, and scheduling latency. In this work, we revisit the pipeline scheduling problem from a principled optimization perspective.\nWe observe that prevailing strategies either rely on static rules or aggressively offload activations without fully leveraging the interaction between memory constraints and scheduling efficiency. To address this, we formulate scheduling as a constrained optimization problem that jointly accounts for memory capacity, activation reuse, and pipeline bubble minimization.\nSolving this model yields fine-grained schedules that reduce pipeline bubbles while adhering to strict memory budgets. Our approach complements existing offloading techniques: whereas prior approaches trade memory for time in a fixed pattern, we dynamically optimize the tradeoff with respect to model structure and hardware configuration.\nExperimental results demonstrate that our method consistently improves both throughput and memory utilization. In particular, we reduce idle pipeline time by up to 50% under the same per-device memory limit, and in some cases, enable the training of larger models within limited memory budgets."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Pipeline Parallelism",
"Scheduling",
"Offloading",
"LLM Training"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/eb86a344ab8ad7fa5d18a9b85971432726eb2c2d.pdf"
},
"primary_area": {
"value": "infrastructure, software libraries, hardware, systems, etc."
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "OptPipe: Memory- and Scheduling-Optimized Pipeline Parallelism for LLM Training"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0A2rXt5SAy",
"id": "0A2rXt5SAy",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission12551/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897502272,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission12551/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission12551/Authors"
]
} | |
2,026 | 0A3qzLmRHd | [
4,
2,
4
] | [
{
"content": "The paper introduces SECVULEVAL, a new vulnerability detection dataset focusing on LLM-based solutions and C/C++ projects. The authors collected the vulnerability data from the national vulnerability database (NVD), and extracted line-level vulnerability labels. Furthermore, they used an LLM to ex... | {
"cdate": 1758230106795,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025secvuleval,\ntitle={{SECVULEVAL}: Benchmarking {LLM}s for Real-World C/C++ Vulnerability Detection},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0A3qzLmRHd},\nnote={under review}\n}"
},
"abstract": {
"value": "Large Language Models (LLMs) have shown promise in various software engineering tasks, but evaluating their effectiveness in vulnerability detection remains challenging due to the lack of high-quality benchmark datasets. Most existing datasets are limited to function-level labels, ignoring finer-grained vulnerability patterns and crucial contextual information. They also often suffer from poor data quality, such as mislabeling, inconsistent annotations, and duplicates, which can lead to inflated performance and weak generalization. Moreover, by including only the vulnerable functions, these datasets miss broader program context, like data/control dependencies and interprocedural interactions, that are essential for accurately detecting and understanding real-world security flaws. Without this context, detection models are evaluated under unrealistic assumptions, limiting their practical impact. To address these limitations, this paper introduces SECVULEVAL, a comprehensive benchmark designed to support fine-grained evaluation of LLMs and other detection methods with rich contextual information. SECVULEVAL focuses on real-world C/C++ vulnerabilities at the statement level. This granularity enables more precise evaluation of a model’s ability to localize and understand vulnerabilities, beyond simple binary classification at the function level. By incorporating rich contextual information, SECVULEVAL sets a new standard for benchmarking vulnerability detection in realistic software development scenarios. This benchmark includes 25,440 function samples covering 5,867 unique CVEs in C/C++ projects from 1999 to 2024. We evaluated the SOTA LLMs with a multi-agent-based approach. The evaluation on our dataset shows that the models are still far from accurately predicting vulnerable statements in a given function. The best-performing Claude-3.7-Sonnet model achieves a 23.83% F1-score for detecting vulnerable statements with correct reasoning, with GPT-4.1 closely behind. We also evaluate the effect of using contextual information for the vulnerability detection task. Finally, we analyze the LLM outputs and provide insights into their behavior in vulnerability detection for C/C++."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Benchmark",
"Security Vunerability",
"Large Language Model"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/9568b0a1094b7803769e4cfc3d795afafd2e0f92.pdf"
},
"primary_area": {
"value": "datasets and benchmarks"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/89003334187309cbc3769306208e64115cd2fa03.zip"
},
"title": {
"value": "SECVULEVAL: Benchmarking LLMs for Real-World C/C++ Vulnerability Detection"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0A3qzLmRHd",
"id": "0A3qzLmRHd",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission14192/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897384880,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission14192/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission14192/Authors"
]
} | |
2,026 | 0A4Uf88pog | [
4,
4,
6
] | [
{
"content": "This paper introduces VERINA (Verifiable Code Generation Arena), a benchmark comprising 189 manually curated programming tasks in Lean for evaluating end-to-end verifiable code generation. The benchmark assesses three foundational tasks—code generation (CodeGen), specification generation (SpecGen)... | {
"cdate": 1758225274549,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025verina,\ntitle={{VERINA}: Benchmarking Verifiable Code Generation},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0A4Uf88pog},\nnote={under review}\n}"
},
"abstract": {
"value": "Large language models (LLMs) are increasingly integrated in software development, but ensuring correctness in LLM-generated code remains challenging and often requires costly manual review. Verifiable code generation---jointly generating code, specifications, and proofs of code-specification alignment---offers a promising path to address this limitation and further unleash LLMs' benefits in coding. Yet, there exists a significant gap in evaluation: current benchmarks often focus on only individual components rather than providing a holistic evaluation framework of all tasks. In this paper, we introduce VERINA (Verifiable Code Generation Arena), a high-quality benchmark enabling a comprehensive and modular evaluation of code, specification, and proof generation as well as their compositions. VERINA consists of 189 manually curated coding tasks in Lean, with detailed problem descriptions, reference implementations, formal specifications, and extensive test suites. Our extensive evaluation of state-of-the-art LLMs reveals significant challenges in verifiable code generation, especially in proof generation, underscoring the need for improving LLM-based theorem provers in verification domains.\nThe best model, OpenAI o4-mini, achieves a 61.4% code correctness rate, 51.0% for specification soundness and completeness, and a mere 3.6% proof success rate (based on one trial per task).\nWe hope VERINA will catalyze progress in verifiable code generation by providing a rigorous and comprehensive benchmark."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"code generation",
"formal verification",
"verifiable code generation",
"AI for math",
"theorem proving",
"AI for code"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/4006dc82db471198ed3d9f554b01e5c1b118c85d.pdf"
},
"primary_area": {
"value": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/8a4f0137ad2c8b021a2908eea1a659d6e34d6d0f.zip"
},
"title": {
"value": "VERINA: Benchmarking Verifiable Code Generation"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0A4Uf88pog",
"id": "0A4Uf88pog",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission13925/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897403022,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission13925/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission13925/Authors"
]
} | |
2,026 | 0A4iQqwwLG | [
2,
4,
4,
4
] | [
{
"content": "The paper presents a new framework to improve long-form video understanding, by incorporating a SLM-based answer clue generation that complements query-based retrieval. The system also incorporates a compressor to summarize frame features into compact tokens, reducing token load while maintaining ... | {
"cdate": 1758336799281,
"content": {
"TLDR": {
"value": "We propose ClueVQA, a novel retrieval framework enhances query-based frame retrieval for VideoQA by generating and integrating supplementary answer clues, leading to improved performance across long-form video benchmarks and various VideoLLMs."
},
"_bibtex": {
"value": "@misc{\nfozilov2025cluevqa,\ntitle={Clue{VQA}: Enhancing Query Based Retrieval in Video-{LLM}s with Answer Clues},\nauthor={Eldor Fozilov and Donggyu Lee and Seokhoon Jeong and Taehwan Kim},\nyear={2025},\nurl={https://openreview.net/forum?id=0A4iQqwwLG}\n}"
},
"abstract": {
"value": "Video-language models have achieved notable success in understanding complex visual narratives and answering fine-grained questions about video content. However, the computational burden of processing long videos - coupled with the growing size of modern models - restricts most approaches to processing only a limited number of frames. A widely adopted strategy to address this limitation is query-based frame retrieval, where frames are selected based on their semantic similarity to the given query. While effective in many cases, such methods are primarily limited to surface-level relevance matching and can fail when faced with implicit, ambiguous, or reasoning-intensive queries, potentially overlooking critical evidence in the video. In this work, we introduce ClueVQA, a novel retrieval framework that improves upon a standard query-based approach by generating and integrating supplementary answer clues and effectively utilizing them for frame selection. The answer clues are derived from the input query and a global scan of the video, which are then used to produce a secondary scoring distribution over frames. This clue-based distribution is then fused with the original query-based frame score distribution to yield a more informed frame selection. The final selected frames are passed to an off-the-shelf Video-LLM for answer generation. Extensive experiments on long-form VideoQA benchmarks, including MLVU, LongVideoBench, and VideoMME, show that our method considerably improves performance over a standard query-based retrieval method across different Video-LLMs."
},
"anonymous_url": null,
"authorids": {
"value": [
"~Eldor_Fozilov1",
"~Donggyu_Lee4",
"~Seokhoon_Jeong2",
"~Taehwan_Kim1"
]
},
"authors": {
"value": [
"Eldor Fozilov",
"Donggyu Lee",
"Seokhoon Jeong",
"Taehwan Kim"
]
},
"code_of_ethics": null,
"keywords": {
"value": [
"video-language models",
"frame selection",
"long video question answering"
]
},
"no_acknowledgement_section": null,
"paperhash": {
"value": "fozilov|cluevqa_enhancing_query_based_retrieval_in_videollms_with_answer_clues"
},
"pdf": {
"value": "/pdf/42125810676c44f8f09af5cce1b0770930fd2869.pdf"
},
"primary_area": {
"value": "applications to computer vision, audio, language, and other modalities"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "ClueVQA: Enhancing Query Based Retrieval in Video-LLMs with Answer Clues"
},
"venue": {
"value": "ICLR 2026 Conference Withdrawn Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Withdrawn_Submission"
}
},
"forum": "0A4iQqwwLG",
"id": "0A4iQqwwLG",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission22893/-/Full_Submission",
"ICLR.cc/2026/Conference/-/Withdrawn_Submission"
],
"license": "CC BY 4.0",
"mdate": 1763019938697,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission22893/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission22893/Authors"
]
} | |
2,026 | 0ACUx9pMWJ | [
6,
4,
4,
6
] | [
{
"content": "The authors propose to study the ability of execution-guided program synthesis approach and transduction approaches with test-time training to generalize to new ARC-AGI-like tasks at test time. Train and test tasks are designed by hand to involve different compositions of the same set of predefine... | {
"cdate": 1758130722525,
"content": {
"TLDR": {
"value": "Comparing the OOD generalization performance of execution-guided neural program synthesis with test-time fine-tuning on the ARC-AGI domain"
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025outofdistribution,\ntitle={Out-of-Distribution Generalization in the {ARC}-{AGI} Domain: Comparing Execution-Guided Neural Program Synthesis and Test-Time Fine-Tuning},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0ACUx9pMWJ},\nnote={under review}\n}"
},
"abstract": {
"value": "We run a controlled compositional generalization experiment in the ARC-AGI domain: an open-world problem domain in which the ability to generalize out-of-distribution is, by design, an essential characteristic for success. We compare neural program synthesis and test-time fine-tuning approaches on this experiment. We find that execution-guided neural program synthesis outperforms all reference algorithms in its ability to compose novel solutions. Our empirical findings also suggest that the success of TTFT on ARC-AGI lies mainly in eliciting in-distribution knowledge that the LLM otherwise fails to rely on directly."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"ARC-AGI",
"Neural program synthesis",
"test-time fine-tuning"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/4edd49acd88859d69d21513a344c02a728179996.pdf"
},
"primary_area": {
"value": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/066fe09216496bc466dccf7dd77ae7519d72640a.zip"
},
"title": {
"value": "Out-of-Distribution Generalization in the ARC-AGI Domain: Comparing Execution-Guided Neural Program Synthesis and Test-Time Fine-Tuning"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0ACUx9pMWJ",
"id": "0ACUx9pMWJ",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission9620/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897708604,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission9620/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission9620/Authors"
]
} | |
2,026 | 0Af7UiJISU | [
6,
6,
4
] | [
{
"content": "THOR is a technically competent and empirically thorough paper that explores hierarchical optimization for tool-integrated reasoning—a topic of rising importance in post-RLHF LLM training. The paper’s main contributions, especially the combination of TIR data generation (TIRGen) and dual-level GRP... | {
"cdate": 1757839164924,
"content": {
"TLDR": {
"value": "We introduce THOR, a tool-integrated framework that combines hierarchical reinforcement learning with self-correcting inference to achieve SOTA mathematical reasoning."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025thor,\ntitle={{THOR}: Tool-Integrated Hierarchical Optimization via {RL} for Mathematical Reasoning},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0Af7UiJISU},\nnote={under review}\n}"
},
"abstract": {
"value": "Large Language Models (LLMs) have made remarkable progress in mathematical reasoning, but still continue to struggle with high-precision tasks like numerical computation and formal symbolic manipulation. Integrating external tools has emerged as a promising approach to bridge this gap. Despite recent advances, existing methods struggle with three key challenges: constructing tool-integrated reasoning data, performing fine-grained optimization, and enhancing inference. To overcome these limitations, we propose THOR (Tool-Integrated Hierarchical Optimization via RL). First, we introduce TIRGen, a multi-agent actor-critic-based pipeline for constructing high-quality datasets of tool-integrated reasoning paths, aligning with the policy and generalizing well across diverse models. Second, to perform fine-grained hierarchical optimization, we introduce an RL strategy that jointly optimizes for both trajectory-level problem solving and step-level code generation. This is motivated by our key insight that the success of an intermediate tool call is a strong predictor of the final answer's correctness. Finally, THOR incorporates a self-correction mechanism that leverages immediate tool feedback to dynamically revise erroneous reasoning paths during inference. Our approach demonstrates strong generalization across diverse models, performing effectively in both reasoning and non-reasoning models. It further achieves state-of-the-art performance for models of a similar scale on multiple mathematical benchmarks, while also delivering consistent improvements on code benchmarks."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Large Language Models",
"Mathematical Problem Solving",
"Tool-Integrated Reasoning",
"Reinforcement Learning"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/1d81b28f8346083153cf34d0bc2540521a6e87bc.pdf"
},
"primary_area": {
"value": "other topics in machine learning (i.e., none of the above)"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/c35d11b552e4f7a5d42cf66249ee8cc24c0071c6.zip"
},
"title": {
"value": "THOR: Tool-Integrated Hierarchical Optimization via RL for Mathematical Reasoning"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0Af7UiJISU",
"id": "0Af7UiJISU",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission5050/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897997957,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission5050/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission5050/Authors"
]
} | |
2,026 | 0B5K9pIdSK | [
4,
6,
2,
6
] | [
{
"content": "This paper introduces TITOK, a framework for transferring LoRA adapters between large language models through token-level contrastive knowledge transfer.\nUnlike existing methods such as TransLoRA, which rely on synthetic data filtered by an external discriminator, TITOK uses a self-contained cont... | {
"cdate": 1758360984163,
"content": {
"TLDR": {
"value": "We propose a new framework TiTok, which enables effective LoRA transplantation through token-level knowledge transfer"
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025titok,\ntitle={TiTok: Transfer Token-level Knowledge via Contrastive Excess to Transplant Lo{RA}},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0B5K9pIdSK},\nnote={under review}\n}"
},
"abstract": {
"value": "Large Language Models (LLMs) are widely applied in real world scenarios, but fine-tuning them comes with significant computational and storage costs. Parameter-Efficient Fine-Tuning (PEFT) methods such as LoRA mitigate these costs, but the adapted parameters are dependent on the base model and cannot be transferred across different backbones. One way to address this issue is through knowledge distillation, but its effectiveness inherently depends on training data. Recent work such as TransLoRA avoids this by generating synthetic data, but this adds complexity because it requires training an additional discriminator model. In this paper, we propose TiTok, a new framework that enables effective LoRA Transplantation through Token-level knowledge transfer. Specifically, TiTok captures task-relevant information through a contrastive excess between a source model with and without LoRA. This excess highlights informative tokens and enables selective filtering of synthetic data, all without additional models or overhead. Through experiments on three benchmarks across multiple transfer settings, our experiments show that the proposed method is consistently effective, achieving average performance gains of +4–8% compared to baselines overall."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Large Language Models",
"Knowledge Transfer",
"PEFT"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/36dec6a7418f68bdb0d5b16bf31ccf444eadf348.pdf"
},
"primary_area": {
"value": "foundation or frontier models, including LLMs"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "TiTok: Transfer Token-level Knowledge via Contrastive Excess to Transplant LoRA"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0B5K9pIdSK",
"id": "0B5K9pIdSK",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission24843/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759896745854,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission24843/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission24843/Authors"
]
} | |
2,026 | 0BD2dCM4Ig | [
2,
2,
6,
2
] | [
{
"content": "This paper studies graph foundation model, and proposes a MoE module together with a graph self-supervised learning-based regularization term to enhance the performance.",
"id": "OopTbQalPe",
"rating": 2
},
{
"content": "The paper studies graph foundation models (GFM) through the l... | {
"cdate": 1757136379041,
"content": {
"TLDR": {
"value": "We propose MoT to address optimization pitfalls in GFMs and achieve SOTA cross-domain generalization."
},
"_bibtex": {
"value": "@misc{\nli2025two,\ntitle={Two Sides of the Same Optimization Coin: Model Degradation and Representation Collapse in Graph Foundation Models},\nauthor={Xunkai Li and Daohan Su and Sicheng Liu and Ru Zhang and Zhenjun Li and Bing Zhou and Rong-Hua Li and Guoren Wang},\nyear={2025},\nurl={https://openreview.net/forum?id=0BD2dCM4Ig}\n}"
},
"abstract": {
"value": "Graph foundation models (GFMs), inspired by the success of LLMs, are designed to learn the optimal embedding function from multi-domain text-attributed graphs (pre-training) for the downstream cross-task generalization capability (fine-tuning). During our investigation, graph vector quantized-masked autoencoder (gVQ-MAE) stands out among the increasingly diverse landscape of GFM architectures. This is attributed to its ability to jointly encode topology and textual attributes from multiple domains into discrete embedding spaces with clear semantic boundaries. Despite its potential, domain generalization conflicts cause imperceptible pitfalls. In this paper, we instantiate two of them, and they are just like two sides of the same GFM optimization coin - Side 1 Model Degradation: The encoder and codebook fail to capture the diversity of inputs (e.g., social networks and molecular graphs); Side 2 Representation Collapse: The hidden embedding and codebook vector fail to preserve semantic separability due to constraints from narrow representation subspaces. These two pitfalls (sides) collectively impair the decoder and generate the low-quality reconstructed supervision, causing the GFM optimization dilemma during pre-training (coin). Through empirical investigation, we attribute the above challenges to Information Bottleneck and Regularization Deficit. To address them, we propose MoT (Mixture-of-Tinkers) - (1) Information Tinker for Two Pitfalls, which utilizes an edge-wise semantic fusion strategy and a mixture-of-codebooks with domain-aware routing to improve information capacity. (2) Regularization Tinker for Optimization Coin, which utilizes two additional regularizations to further improve gradient supervision in our proposed Information Tinker. Notably, as a flexible architecture, MoT adheres to the scaling laws of GFM, offering a controllable model scale. Compared to SOTA baselines, experiments on 22 datasets across 6 domains demonstrate that MoT achieves significant improvements in supervised (1.4%), few-shot (3.1%), and zero-shot (3.3%) scenarios."
},
"anonymous_url": null,
"authorids": {
"value": [
"~Xunkai_Li1",
"~Daohan_Su1",
"~Sicheng_Liu2",
"~Ru_Zhang8",
"~Zhenjun_Li1",
"~Bing_Zhou3",
"~Rong-Hua_Li2",
"~Guoren_Wang1"
]
},
"authors": {
"value": [
"Xunkai Li",
"Daohan Su",
"Sicheng Liu",
"Ru Zhang",
"Zhenjun Li",
"Bing Zhou",
"Rong-Hua Li",
"Guoren Wang"
]
},
"code_of_ethics": null,
"keywords": {
"value": [
"Foundation Model; Graph Pre-training; Vector Quantization"
]
},
"no_acknowledgement_section": null,
"paperhash": {
"value": "li|two_sides_of_the_same_optimization_coin_model_degradation_and_representation_collapse_in_graph_foundation_models"
},
"pdf": {
"value": "/pdf/2abdb6e4ce624e14ff91cd45148798851f75c236.pdf"
},
"primary_area": {
"value": "foundation or frontier models, including LLMs"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/fedf3aa3843901c06791f68b4f82d7bafef5c686.zip"
},
"title": {
"value": "Two Sides of the Same Optimization Coin: Model Degradation and Representation Collapse in Graph Foundation Models"
},
"venue": {
"value": "ICLR 2026 Conference Withdrawn Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Withdrawn_Submission"
}
},
"forum": "0BD2dCM4Ig",
"id": "0BD2dCM4Ig",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/-/Withdrawn_Submission"
],
"license": "CC BY 4.0",
"mdate": 1763275020400,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission2528/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission2528/Authors"
]
} | |
2,026 | 0BWu7DLuIU | [
2,
2,
4,
2
] | [
{
"content": "This paper considers the ethical implications of using Homomorphic Encryption (HE). Specifically, they investigate the desirable outcomes (The Good), trade-offs related to accountability, interpretability and responsibility (The Bad) and ways in which HE can be used to mask unethical practices (Th... | {
"cdate": 1757316151995,
"content": {
"TLDR": {
"value": "Privacy is not everything in privacy-preserving Artificial Intelligence with Homomorphic Encryption."
},
"_bibtex": {
"value": "@misc{\nfalcetta2025the,\ntitle={The Ethics of Privacy-Preserving Deep Learning: the Good, the Bad, and the Ugly},\nauthor={Alessandro Falcetta and Stefano Canali and Viola Schiaffonati and Manuel Roveri},\nyear={2025},\nurl={https://openreview.net/forum?id=0BWu7DLuIU}\n}"
},
"abstract": {
"value": "Homomorphic Encryption (HE) is gaining traction in Artificial Intelligence (AI) as a solution to privacy concerns, particularly in sensitive areas such as health-care. HE enables computation directly on encrypted data without ever decrypting it, ensuring that private information remains protected throughout the process, effectively enabling Privacy-Preserving Machine and Deep Learning (PP-MDL). While much of the discussion focuses on its privacy benefits, little attention has been paid to the ethical implications of using and training AI models on encrypted data, especially when the resulting models themselves remain encrypted and opaque to both developers and users. In this paper, we explore three ethical perspectives on the use of HE for PP-MDL: the Good, i.e., the clear advantages in terms of privacy and data protection; the Bad, i.e., the practical and conceptual ethical challenges it introduces; and the Ugly, i.e., the subtle, unexpected ethical issues that may arise when HE-powered PP-MDL is deployed in the real-world. Our aim is to show that while HE can strengthen privacy, it is not a silver bullet for ethical AI. It can complicate accountability, transparency, and trust, raising important ethical and societal questions that should not be overlooked."
},
"anonymous_url": null,
"authorids": {
"value": [
"~Alessandro_Falcetta1",
"~Stefano_Canali1",
"~Viola_Schiaffonati1",
"~Manuel_Roveri2"
]
},
"authors": {
"value": [
"Alessandro Falcetta",
"Stefano Canali",
"Viola Schiaffonati",
"Manuel Roveri"
]
},
"code_of_ethics": null,
"keywords": {
"value": [
"artificial intelligence",
"homomorphic encryption",
"ethics",
"privacy"
]
},
"no_acknowledgement_section": null,
"paperhash": {
"value": "falcetta|the_ethics_of_privacypreserving_deep_learning_the_good_the_bad_and_the_ugly"
},
"pdf": {
"value": "/pdf/205efb398f7dea966a0d6bd7b731332c2cbb6d1d.pdf"
},
"primary_area": {
"value": "alignment, fairness, safety, privacy, and societal considerations"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "The Ethics of Privacy-Preserving Deep Learning: the Good, the Bad, and the Ugly"
},
"venue": {
"value": "ICLR 2026 Conference Withdrawn Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Withdrawn_Submission"
}
},
"forum": "0BWu7DLuIU",
"id": "0BWu7DLuIU",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission3011/-/Full_Submission",
"ICLR.cc/2026/Conference/-/Withdrawn_Submission"
],
"license": "CC BY 4.0",
"mdate": 1762966891512,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission3011/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission3011/Authors"
]
} | |
2,026 | 0BhjNjxpaC | [
2,
8,
8,
2
] | [
{
"content": "This work studies the question: \"how many 'reasoning steps' can an $L$-layer Transformer carry out in a single forward pass?\". To answer this question, the paper posits a formulation of \"reasoning chains\" as sequences of pairs of integers $a_i^1 \\to a_i^2$, which can be permuted. It then form... | {
"cdate": 1757819222345,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025limit,\ntitle={Limit Analysis for Symbolic Multi-step Reasoning Tasks with Information Propagation Rules Based on Transformers},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0BhjNjxpaC},\nnote={under review}\n}"
},
"abstract": {
"value": "Transformers have ability to perform reasoning tasks, however the intrinsic mechanism remains widely open. In this paper we propose a set of information propagation rules based on Transformers and utilize symbolic reasoning tasks to theoretically analyze the limit reasoning steps. \nWe show that the number of reasoning steps has an upper bound of $O(3^{L-1})$ and a lower bound of $O(2^{L-1})$ for a model with $L$ attention layers."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"multi-step reasoning",
"parallel reasoning",
"large language models",
"interpretability",
"buffer mechanism",
"model capacity"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/f02231666f3fcc43712aa327cc31312d8550481e.pdf"
},
"primary_area": {
"value": "learning theory"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Limit Analysis for Symbolic Multi-step Reasoning Tasks with Information Propagation Rules Based on Transformers"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0BhjNjxpaC",
"id": "0BhjNjxpaC",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission4954/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759898003047,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission4954/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission4954/Authors"
]
} | |
2,026 | 0BkvUY61MX | [
6,
2,
8
] | [
{
"content": "1. The paper presents scaling laws in the multilingual setup across different axis:\n\n1.1 For repeated epochs in the monolingual setup\n\n1.2 To account for cross lingual transfer in a language mixture setup\n\n1.3. Account for the curse of multilinguality by providing a closed form approximation... | {
"cdate": 1758341769702,
"content": {
"TLDR": {
"value": "Scaling laws for multilingual pretraining, finetuning, and language transfer."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025atlas,\ntitle={{ATLAS}: Adaptive Transfer Scaling Laws for Multilingual Pretraining, Finetuning, and Decoding the Curse of Multilinguality},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0BkvUY61MX},\nnote={under review}\n}"
},
"abstract": {
"value": "Scaling laws research has focused overwhelmingly on English—yet the most prominent AI models explicitly serve billions of international users. In this work, we undertake the largest multilingual scaling laws study to date, totaling 774 multilingual training experiments, spanning 10M-8B model parameters, 400+ training languages and 48 evaluation languages. We introduce the Adaptive Transfer Scaling Law (ATLAS) for both monolingual and multilingual pretraining, which outperforms existing scaling laws' out-of-sample generalization often by more than 0.3 R². Our analyses of the experiments shed light on multilingual learning dynamics, transfer properties between languages, and the curse of multilinguality. First, we derive a cross-lingual transfer matrix, empirically measuring mutual benefit scores between 38 × 38 = 1444 language pairs. Second, we derive a language-agnostic scaling law that reveals how to optimally scale model size and data when adding languages without sacrificing performance. Third, we identify the computational crossover points for when to pretrain from scratch versus finetune from multilingual checkpoints. We hope these findings provide the scientific foundation for democratizing scaling laws across languages, and enable practitioners to efficiently scale models—beyond English-first AI."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"scaling laws",
"multilinguality"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/8058aba8e6566d3587e70df9f688d9784b59a1bf.pdf"
},
"primary_area": {
"value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "ATLAS: Adaptive Transfer Scaling Laws for Multilingual Pretraining, Finetuning, and Decoding the Curse of Multilinguality"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0BkvUY61MX",
"id": "0BkvUY61MX",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission23289/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759896822547,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission23289/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission23289/Authors"
]
} | |
2,026 | 0CQnhxpE7w | [
4,
8,
4,
6
] | [
{
"content": "The authors propose EVCtrl, a training free method for speeding up inference of ControlNet based models. The method is based on the observation that when using sparse conditioning modalities mostly consisting of black pixels, only tokens corresponding to non-black regions are critical for updating... | {
"cdate": 1758260698841,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@misc{\nyang2025evctrl,\ntitle={{EVC}trl: Efficient Control Adapter for Visual Generation},\nauthor={Zixiang Yang and Yue Ma and Yinhan Zhang and Shanhui Mo and Dongrui Liu and Linfeng Zhang},\nyear={2025},\nurl={https://openreview.net/forum?id=0CQnhxpE7w}\n}"
},
"abstract": {
"value": "Visual generation includes both image and video generation, training probabilistic models to create coherent, diverse, and semantically faithful content from scratch. While early research focused on unconditional sampling, practitioners now demand controllable generation that allows precise specification of layout, pose, motion, or style. While ControlNet grants precise spatial-temporal control, its auxiliary branch markedly increases latency and introduces redundant computation in both uncontrolled regions and denoising steps, especially for video. To address this problem, we introduce EVCtrl, a lightweight, plug-and-play control adapter that slashes overhead without retraining the model. Specifically, we propose a spatio-temporal dual caching strategy for sparse control information. For spatial redundancy, we first profile how each layer of DiT-ControlNet responds to fine-grained control, then partition the network into global and local functional zones. A locality-aware cache focuses computation on the local zones that truly need the control signal, skipping the bulk of redundant computation in global regions. For temporal redundancy, we selectively omit unnecessary denoising steps to improve efficiency. Extensive experiments on CogVideo-Controlnet, Wan2.1-Controlnet, and Flux demonstrate that our method is effective in image and video control generation without the need for training. For example, it achieves 2.16 and 2.05 times speedups on CogVideo-Controlnet and Wan2.1-Controlnet, respectively, with almost no degradation in generation quality.Codes are available in the supplementary materials."
},
"anonymous_url": null,
"authorids": {
"value": [
"~Zixiang_Yang1",
"~Yue_Ma2",
"~Yinhan_Zhang1",
"~Shanhui_Mo2",
"~Dongrui_Liu1",
"~Linfeng_Zhang2"
]
},
"authors": {
"value": [
"Zixiang Yang",
"Yue Ma",
"Yinhan Zhang",
"Shanhui Mo",
"Dongrui Liu",
"Linfeng Zhang"
]
},
"code_of_ethics": null,
"keywords": {
"value": [
"Visual Generation",
"Diffusion Models",
"Control Adapter"
]
},
"no_acknowledgement_section": null,
"paperhash": {
"value": "yang|evctrl_efficient_control_adapter_for_visual_generation"
},
"pdf": {
"value": "/pdf/4e4a0c89b75c377756da0a0ce2fd4fac6083971b.pdf"
},
"primary_area": {
"value": "generative models"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/6310dd2067b10a2bd41d3015e0c97ea35f348fad.zip"
},
"title": {
"value": "EVCtrl: Efficient Control Adapter for Visual Generation"
},
"venue": {
"value": "ICLR 2026 Conference Withdrawn Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Withdrawn_Submission"
}
},
"forum": "0CQnhxpE7w",
"id": "0CQnhxpE7w",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission16151/-/Full_Submission",
"ICLR.cc/2026/Conference/-/Withdrawn_Submission"
],
"license": "CC BY 4.0",
"mdate": 1762943364931,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission16151/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission16151/Authors"
]
} | |
2,026 | 0CXjpAxHUE | [
8,
6,
4,
8
] | [
{
"content": "This paper presents a theoretical framework for understanding multi-epoch data reuse in the context of linear regression and its implications for data-scaling laws in large model training. It shows that larger datasets can be repeated more times effectively. Simulation and LLM pretraining experime... | {
"cdate": 1758246906466,
"content": {
"TLDR": {
"value": "Theoretical analysis of multi-epoch scaling in linear regression"
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025larger,\ntitle={Larger Datasets Can Be Repeated More: A Theoretical Analysis of Multi-Epoch Scaling in Linear Regression},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0CXjpAxHUE},\nnote={under review}\n}"
},
"abstract": {
"value": "Large Language Model (LLM) training often processes vast text corpora in a single pass, leaving much available data underutilized. This paper presents a theoretical analysis of how a common workaround, training for multiple epochs on the same dataset, reshapes the data scaling laws. Concretely, given a $K$-epoch training on $N$ samples, how many fresh samples would one-pass training require to match the same performance? We quantify this using the \\textit{effective reuse rate} of the data, $E(K, N)$, which we define as the factor by which the dataset must grow under one-pass training to match the test loss of multi-epoch training. Our analysis precisely characterizes the scaling behavior of $E(K, N)$ for SGD in linear regression under either strong convexity or Zipf-distributed data: (1) When $K$ is small, we prove that $E(K, N) \\approx K$, indicating that every new epoch yields a linear gain; (2) As $K$ increases, $E(K, N)$ plateaus at a problem-dependent value that grows with $N$ ($\\Theta(\\log N)$ for the strongly-convex case), implying that larger datasets can be repeated more times before the marginal benefit vanishes. These theoretical findings complement a recent empirical study by [Muennighoff et al. (2023)](https://arxiv.org/abs/2305.16264), which found that training LLMs for up to $4$ epochs results in negligible loss differences compared to using fresh data at each step, \\textit{i.e.}, $E(K, N) \\approx K$ for $K \\le 4$ in our notation. \n Supported by further empirical validation with LLMs, our results reveal how this behavior depends on the underlying data size and distribution, and underscore the need to explicitly model both factors in future studies of scaling laws with data reuse."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Deep learning theory",
"Multi-epoch training",
"Data-reuse",
"Optimization",
"Scaling law",
"Large language model"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/51ee840632db57dc0e2199f27424cf25620fc996.pdf"
},
"primary_area": {
"value": "learning theory"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Larger Datasets Can Be Repeated More: A Theoretical Analysis of Multi-Epoch Scaling in Linear Regression"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0CXjpAxHUE",
"id": "0CXjpAxHUE",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission15017/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897335276,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission15017/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission15017/Authors"
]
} | |
2,026 | 0CZAimzcVr | [
6,
6,
6,
6
] | [
{
"content": "The paper studies a quite general optimization problem, namely maximizing a function F on [0,1]^n under some contraints defining a feasible set $\\cal C \\in [0,1]^n$. The function is DR-submodular. The feasible set is convex. Such a problem is typically solved using gradient ascend, and there is ... | {
"cdate": 1758106456114,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025drsubmodular,\ntitle={{DR}-Submodular Maximization with Stochastic Biased Gradients: Classical and Quantum Gradient Algorithms},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0CZAimzcVr},\nnote={under review}\n}"
},
"abstract": {
"value": "In this work, we investigate DR-submodular maximization using stochastic biased gradients, which is a more realistic but challenging setting than stochastic unbiased gradients. We first generalize the Lyapunov framework to incorporate biased stochastic gradients, characterizing the adverse impacts of bias and noise. Leveraging this framework, we consider not only conventional constraints but also a novel constraint class: convex sets with a largest element, which naturally arises in applications such as resource allocations. For this constraint, we propose an $1/e$ approximation algorithm for non-monotone DR-submodular maximization, surpassing the hardness result $1/4$ for general convex constraints. As a direct application of stochastic biased gradients, we consider zero-order DR-submodular maximization and introduce both classical and quantum gradient estimation algorithms. In each constraint we consider, while retaining the same approximation ratio, the iteration complexity of our classical zero-order algorithms is $O(\\epsilon^{-3})$, matching that of stochastic unbiased gradients; our quantum zero-order algorithms reach $O(\\epsilon^{-1})$ iteration complexity, on par with classical first-order algorithms, demonstrating quantum acceleration and validated in numerical experiments."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"DR-submodular Maximization",
"Stochastic Biased Gradients",
"Zero-Order Optimization",
"Quantum Gradient Estimation",
"Approximation Algorithms"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/a13ab47d1b8bef974ca2dc3482b61e402ac773a8.pdf"
},
"primary_area": {
"value": "optimization"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/7bc45e059cf0db278334218ed862e2f62c1f0345.pdf"
},
"title": {
"value": "DR-Submodular Maximization with Stochastic Biased Gradients: Classical and Quantum Gradient Algorithms"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0CZAimzcVr",
"id": "0CZAimzcVr",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission8996/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897749083,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission8996/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission8996/Authors"
]
} | |
2,026 | 0CajQNVKyB | [
8,
6,
6,
4
] | [
{
"content": "This paper introduces HERO (Hybrid Ensemble Reward Optimization)- a framework that integrates dense signals from reward models with binary-valued feedback from rule-based verifiers. The paper systematically reports the merits of each individual approach while highlighting its limitations; they fur... | {
"cdate": 1758300838198,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025hybrid,\ntitle={Hybrid Reinforcement: when reward is sparse, better to be dense},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0CajQNVKyB},\nnote={under review}\n}"
},
"abstract": {
"value": "Post-training for reasoning of Large language models (LLMs) increasingly rely on verifiable rewards: deterministic checkers that provide 0–1 correctness signals. While reliable, such binary feedback is brittle—many tasks admit partially correct or alternative answers that verifiers under-credit, and the resulting all-or-nothing supervision limits learning. Reward models offer richer, continuous feedback, which can serve as a complementary supervisory signal to verifiers. We introduce HERO (Hybrid Ensemble Reward Optimization), a reinforcement learning framework that integrates verifier signals with reward-model scores in a structured way. HERO employs stratified normalization to bound reward-model scores within verifier-defined groups, preserving correctness while refining quality distinctions, and variance-aware weighting to emphasize challenging prompts where dense signals matter most. Across diverse mathematical reasoning benchmarks, HERO consistently outperforms RM-only and verifier-only baselines, with strong gains on both verifiable and hard-to-verify tasks. Our results show that hybrid reward design retains the stability of verifiers while leveraging the nuance of reward models to advance reasoning."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Hybrid rewards for reinforcement learning"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/1b9434a429cc2bf09028a5a7aefb7e84e5534b50.pdf"
},
"primary_area": {
"value": "foundation or frontier models, including LLMs"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Hybrid Reinforcement: when reward is sparse, better to be dense"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0CajQNVKyB",
"id": "0CajQNVKyB",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission19944/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897011255,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission19944/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission19944/Authors"
]
} | |
2,026 | 0Cv0whP7l8 | [
6,
4,
4,
6
] | [
{
"content": "This paper introduces a framework to diagnose and mitigate modality interference in multimodal large language models (MLLMs)—a phenomenon where irrelevant or misleading modality signals degrade model performance. The authors define the broader cross-modality competency problem, identifying modalit... | {
"cdate": 1758184011136,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025diagnosing,\ntitle={Diagnosing and Mitigating Modality Interference in Multimodal Large Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0Cv0whP7l8},\nnote={under review}\n}"
},
"abstract": {
"value": "Multimodal Large Language Models have demonstrated impressive capabilities across tasks, yet they often exhibit difficulty in distinguishing task-relevant from irrelevant signals—particularly in tasks like Visual Question Answering—which can lead to susceptibility to misleading or spurious inputs. We refer to this broader limitation as the Cross-Modality Competency Problem—the model’s inability to fairly evaluate all modalities. This vulnerability becomes more evident in modality-specific tasks—such as image classification or pure text question answering—where models are expected to rely solely on one modality. In such tasks, spurious information from irrelevant modalities often leads to significant performance degradation. We refer to this failure as Modality Interference, which serves as a concrete and measurable instance of the cross-modality competency problem, and we further design a perturbation-based causal diagnostic experiment to verify and quantify this problem. To mitigate modality interference, we propose a novel framework to finetune MLLMs, including perturbation-based data augmentations with both heuristic perturbations and adversarial perturbations, and a consistency regularization strategy applying on model outputs with original and perturbed inputs. Experiments on multiple benchmark datasets (image-heavy, text-heavy and multimodal tasks) and multiple model families with different scales demonstrate significant improvements in robustness and cross-modality competency, indicating our method's effectiveness in boosting unimodal reasoning ability while enhancing performance on multimodal tasks."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Multimodal Large Language Models",
"Modality Interference",
"Causal Intervention"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/7a652a4137746f54b609bc9a256c5f5c918f2c97.pdf"
},
"primary_area": {
"value": "foundation or frontier models, including LLMs"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/0c5cfb17d948f2eba0a86516a122f3174a871ac7.zip"
},
"title": {
"value": "Diagnosing and Mitigating Modality Interference in Multimodal Large Language Models"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0Cv0whP7l8",
"id": "0Cv0whP7l8",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission10880/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897622846,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission10880/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission10880/Authors"
]
} | |
2,026 | 0Cv9PwL7cI | [
8,
4,
4,
6
] | [
{
"content": "This paper investigates the limitations of fixed block-size semi-autoregressive decoding in diffusion-based large language models (dLLMs). The authors identify two inefficiencies — Late Decoding Overhead and Premature Decoding Error — that arise when fixed-size blocks fail to align with semantic s... | {
"cdate": 1756733104796,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025semanticaware,\ntitle={Semantic-Aware Diffusion {LLM} Inference With Adaptive Block Size},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0Cv9PwL7cI},\nnote={under review}\n}"
},
"abstract": {
"value": "Diffusion-based large language models (dLLMs) are gaining attention for their inherent capacity for parallel decoding, offering a compelling alternative to autoregressive LLMs. Among various decoding strategies, blockwise semi-autoregressive (semi-AR) approaches are widely adopted due to their natural support for KV caching and their favorable accuracy–speed trade-off. However, this paper identifies two fundamental limitations in the conventional semi-AR decoding approach that applies a fixed block size: i) late decoding overhead, where the unmasking of high-confidence tokens outside the current block is unnecessarily delayed, and ii) premature decoding error, where low-confidence tokens inside the current block are committed too early, leading to incorrect tokens. This paper presents the first systematic investigation challenging the fixed block size assumption in semi-AR decoding. Through a statistical analysis of confidence dynamics during the denoising process, we identify a volatility band (VB) region during dLLM decoding, which encodes local semantic structure and can be used to guide adaptive block sizing. Leveraging these insights, we introduce AdaBlock-dLLM, a training-free, plug-and-play scheduler that adaptively aligns block boundaries with semantic steps by adjusting block size during runtime. Extensive experiments across diverse benchmarks show that AdaBlock-dLLM achieves up to 5.3% accuracy improvement under the same throughput budget. Beyond inference-time optimization, we hope our semantics-aware adaptive scheduling approach and confidence-based analysis will inspire future training strategies for dLLMs."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Diffusion Large Language Models",
"Non-Autoregressive Decoding"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/3d28ee2d83cf87267c69760ca94c762c93cbe5fa.pdf"
},
"primary_area": {
"value": "generative models"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/9f1d83373188a93be3c31dd3a48db5b6b39307e5.zip"
},
"title": {
"value": "Semantic-Aware Diffusion LLM Inference With Adaptive Block Size"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0Cv9PwL7cI",
"id": "0Cv9PwL7cI",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission272/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759898269585,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission272/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission272/Authors"
]
} | |
2,026 | 0DaB4jeGaf | [
4,
6,
4
] | [
{
"content": "This paper proposes a quantile regression framework, where an ReLU neural network is used to approximate the quantile function, and a convolution-type smooth quantile loss is used to train the network. Experimental results on synthetic data show that the proposed framework outperforms ReLU network... | {
"cdate": 1758039348094,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025conquer,\ntitle={Conquer the Quantile: Convolution-Smoothed Quantile Regression with Neural Networks and Minimax Guarantees},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0DaB4jeGaf},\nnote={under review}\n}"
},
"abstract": {
"value": "Quantile regression provides a flexible approach to modeling heterogeneous effects and tail behaviors. This paper introduces the first quantile neural network estimator built upon the \\textbf{con}volution-type smoothing \\textbf{qu}antil\\textbf{e} \\textbf{r}egression (known as \\textit{conquer}) framework, which preserves both convexity and differentiability while retaining the robustness of the quantile loss. Extending the conquer estimator beyond linear models, we develop a nonparametric deep learning framework and establish sharp statistical guarantees. Specifically, we show that our estimator attains the minimax convergence rate over Besov spaces up to a logarithmic factor, matching the fundamental limits of nonparametric quantile estimation, and further derive general upper bounds for the estimation error in more general function classes. Empirical studies demonstrate that our method consistently outperforms existing quantile networks in both estimating accuracy and computational efficiency, underscoring the benefits of incorporating conquer into deep quantile learning."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"quantile regression",
"minimax rate",
"convolution",
"deep learning theory",
"Besov space"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/6472a64f207cbb489dffe7fe14f19ce1bb6da913.pdf"
},
"primary_area": {
"value": "learning theory"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Conquer the Quantile: Convolution-Smoothed Quantile Regression with Neural Networks and Minimax Guarantees"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0DaB4jeGaf",
"id": "0DaB4jeGaf",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission7858/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897826972,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission7858/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission7858/Authors"
]
} | |
2,026 | 0DekoBl3te | [
8,
2,
4,
2
] | [
{
"content": "This paper proposes Dual-MoE, a dual-path mixture-of-experts framework for multivariate time series forecasting that jointly models temporal distribution shifts and noisy channel dependencies. The model consists of two complementary components: the Temporal Fusion MoE and the Channel Fusion MoE. A... | {
"cdate": 1758115541458,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025dualmoe,\ntitle={Dual-MoE: Learning Time and Channel Dependencies via Dual Mixture-of-Experts for Time Series Forecasting},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0DekoBl3te},\nnote={under review}\n}"
},
"abstract": {
"value": "Multivariate time series forecasting holds significant value in finance, energy, and transportation systems, yet faces critical challenges in jointly modeling temporal heterogeneity and dynamic channel dependencies. Existing approaches exhibit limitations in balancing long-term trends with short-term fluctuations, while struggling to capture time-varying inter-variable relationships. This paper proposes Dual-MoE, a dual mixture-of-experts framework that synergistically integrates temporal and channel modeling. The temporal expert dynamically combines multi-scale historical features (e.g., hourly details and weekly patterns) through adaptive gating mechanisms, whereas the channel expert learns dependency weights between variables via frequency-aware interaction modeling. Extensive experiments on real-world datasets demonstrate Dual-MoE's superior forecasting accuracy and robustness compared to state-of-the-art baselines. Its modular architecture provides a flexible and scalable paradigm for complex temporal dependency modeling, paving the way for further advancements in time series analysis. Code is available in Appendix."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Time Series Forecasting"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/5d91d2446dc98942c3a8851d632fb64f3af14172.pdf"
},
"primary_area": {
"value": "learning on time series and dynamical systems"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/5f14436e934acc36691aa0e8867c141ea4f6105d.zip"
},
"title": {
"value": "Dual-MoE: Learning Time and Channel Dependencies via Dual Mixture-of-Experts for Time Series Forecasting"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0DekoBl3te",
"id": "0DekoBl3te",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission9221/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897737118,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission9221/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission9221/Authors"
]
} | |
2,026 | 0Dhpt9aY3n | [
6,
8,
2,
6
] | [
{
"content": "This paper introduces DeepSynth, a very challenging benchmark for evaluating LLM agents. DeepSynth consists of 120 diverse tasks created by 16 experts, where each task requires an agent to navigate through about 4 web pages and read up to 15 documents and tables. The tasks are designed to reflect ... | {
"cdate": 1758304923126,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025a,\ntitle={A Benchmark for Deep Information Synthesis},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0Dhpt9aY3n},\nnote={under review}\n}"
},
"abstract": {
"value": "Large language model (LLM)-based agents are increasingly used to solve complex tasks involving tool use, such as web browsing, code execution, and data analysis. However, current evaluation benchmarks do not adequately assess their ability to solve real-world tasks that require synthesizing information from multiple sources and inferring insights beyond simple fact retrieval. To address this, we introduce DEEPSYNTH , a novel benchmark designed to evaluate agents on realistic, time-consuming problems that combine information gathering, synthesis, and structured reasoning to produce insights. DEEPSYNTH contains 120 tasks collected across 7 domains and data sources covering 42 countries. DEEPSYNTH is constructed using a multi-stage data collection pipeline that requires annotators to collect official data sources, create hypotheses, perform manual analysis and design tasks with verifiable answers. When evaluated on DEEPSYNTH, 9 state-of-the-art LLMs and deep research agents achieve a maximum F1 score of 8.97. Our analysis reveals that current agents struggle with hallucinations and reasoning over large information spaces, highlighting \\ourdata as a crucial benchmark for guiding future research."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Benchmark",
"Deep Information Synthesis",
"LLM agents",
"Deep Research",
"AI agents"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/6b0392289d2433b7468aed713a9794645dc21cad.pdf"
},
"primary_area": {
"value": "datasets and benchmarks"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "A Benchmark for Deep Information Synthesis"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0Dhpt9aY3n",
"id": "0Dhpt9aY3n",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission20343/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759896982619,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission20343/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission20343/Authors"
]
} | |
2,026 | 0DqB1vxGTn | [
2,
4,
2,
6
] | [
{
"content": "The paper propose a method to train a model that predicts depth map from a single image at metric scale. Real-world camera heights are assumed to be known during training and is used to recover metric depths, which are then used as pseudo label ground-truth depth to supervise another student netwo... | {
"cdate": 1757487648325,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@misc{\nzhang2025enhancing,\ntitle={Enhancing Self-Supervised Depth Estimation Through Camera Parameter Priors},\nauthor={Jinchang Zhang and Xue Iuan Wong and Guoyu Lu},\nyear={2025},\nurl={https://openreview.net/forum?id=0DqB1vxGTn}\n}"
},
"abstract": {
"value": "Depth estimation is a key topic in the field of computer vision. Self-supervised monocular depth estimation offers a powerful method to extract 3D scene information from a single camera image, allowing training on arbitrary image sequences without the need for depth labels. However, monocular unsupervised depth estimation still cannot address the issue of scale and often requires ground-truth depth data for calibration.\nIn the deep learning era, existing methods primarily rely on relationships between images to train unsupervised neural networks, often overlooking the foundational information provided by the camera itself. In fact, based on physical principles, the camera’s intrinsic and extrinsic parameters can be used to calculate depth information for the ground and related areas and extend it from planar regions to full scene depth. To make full use of scene depth, even in the presence of errors, we introduce a contrastive learning self-supervised framework. This framework consists of two networks with the same structure: the Anchor network and the Target network. The predictions from the Anchor network are used as pseudo-labels for training the Target network. Depth reliability is determined by entropy, dividing the predicted depth into positive and negative samples to maximize the use of physical depth information, and effectively enhance the depth estimation accuracy."
},
"anonymous_url": null,
"authorids": {
"value": [
"~Jinchang_Zhang2",
"~Xue_Iuan_Wong1",
"~Guoyu_Lu4"
]
},
"authors": {
"value": [
"Jinchang Zhang",
"Xue Iuan Wong",
"Guoyu Lu"
]
},
"code_of_ethics": null,
"keywords": {
"value": [
"Depth Estimation",
"Camera"
]
},
"no_acknowledgement_section": null,
"paperhash": {
"value": "zhang|enhancing_selfsupervised_depth_estimation_through_camera_parameter_priors"
},
"pdf": {
"value": "/pdf/9f05a7e06ca3f5d4f64cb8fa0677eca60d89ae49.pdf"
},
"primary_area": {
"value": "applications to computer vision, audio, language, and other modalities"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/13a7cb39cd35a6f472af021230d2c0ea29633c59.pdf"
},
"title": {
"value": "Enhancing Self-Supervised Depth Estimation Through Camera Parameter Priors"
},
"venue": {
"value": "ICLR 2026 Conference Withdrawn Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Withdrawn_Submission"
}
},
"forum": "0DqB1vxGTn",
"id": "0DqB1vxGTn",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission3621/-/Full_Submission",
"ICLR.cc/2026/Conference/-/Withdrawn_Submission"
],
"license": "CC BY 4.0",
"mdate": 1763330917332,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission3621/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission3621/Authors"
]
} | |
2,026 | 0EV92jgJaZ | [
2,
2,
6
] | [
{
"content": "Knowledge compilation approaches in probabilistic answer set programming (PASP) can be categorised into top-down or bottom-up approaches.\nTop-down typically require a CNF as input, which generally means additional auxiliary variables are introduced to first transform the PASP program into a CNF.\... | {
"cdate": 1758322135358,
"content": {
"TLDR": {
"value": "We propose a non-incremental approach for Bottom-Up Knowledge Compilation of Probabilistic Answer Set programs."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025nonincremental,\ntitle={Non-Incremental Bottom-Up Knowledge Compilation of Neuro-Answer Set Programs},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0EV92jgJaZ},\nnote={under review}\n}"
},
"abstract": {
"value": "Neuro-Probabilistic Answer Set Programming offers an intuitive and expressive framework for representing knowledge involving relations, non-determinism, logical constraints, and uncertainty-aware perception. Such a high expressivity comes at a significant computational cost. To mitigate that, Knowledge Compilation (KC) approaches translate the logic program into a logic circuit for which inference and learning can be performed efficiently. Top-down KC approaches employ an intermediary step of translating the logic program into a CNF propositional formula, before the actual KC step. This has the drawback of requiring the use of auxiliary variables and a fixed variable ordering. Bottom-up KC approaches instead construct a circuit representation compositionally, by employing circuit operations that represent the subparts of the logic program, without the need of auxiliary variables and allowing dynamic variable ordering. However, intermediary circuits can grow quite large even when the end circuit is succinct. In this work, we develop a non-incremental bottom-up KC strategy that provably and empirically reduces the size of the intermediary representations compared to its incremental counterpart. We explore heuristics for v-tree initialization and dynamic variable reordering. Experimental results show that our method achieves state-of-the-art performance for a large class of programs."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Probabilistic Logic Programming",
"Knowledge Compilation"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/8c15560d7d1761b796b890a6ed9a9cbd6b6b139a.pdf"
},
"primary_area": {
"value": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/a76f1ad33861acd68d7a78ec7105bdd343f6ef04.zip"
},
"title": {
"value": "Non-Incremental Bottom-Up Knowledge Compilation of Neuro-Answer Set Programs"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0EV92jgJaZ",
"id": "0EV92jgJaZ",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission21809/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759896902012,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission21809/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission21809/Authors"
]
} | |
2,026 | 0EXuliYnfW | [
4,
2,
4,
6
] | [
{
"content": "This paper propose PPBoost , a method proposed to tackle zero-shot medical image segmentation by bridging the gap between text prompts and visual prompts. PPBoost progressively transform a natural language description of the target anatomy into a high-quality spatial bounding box, which then guide... | {
"cdate": 1758295172014,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025ppboost,\ntitle={{PPBOOST}: {PROGRESSIVE} {PROMPT} {BOOSTING} {FOR} {TEXT}-{DRIVEN} {MEDICAL} {IMAGE} {SEGMENTATION}},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0EXuliYnfW},\nnote={under review}\n}"
},
"abstract": {
"value": "Text-prompted foundation models for medical image segmentation offer an intuitive\nway to delineate anatomical structures from natural language queries, but\ntheir predictions often lack spatial precision and degrade under domain shift.\nIn contrast, visual-prompted models achieve strong segmentation performance\nacross diverse modalities by leveraging spatial cues of precise bounding-box\n(bbox) prompts to guide the segmentation of target lesions. However, it is costly\nand challenging to obtain the precise visual prompts in clinical practice. We propose\nPPBoost (Progressive Prompt-Boosting), a framework that bridges these limitations\nby transforming weak text-derived signals into strong, spatially grounded\nvisual prompts, operating under a strict zero-shot regime with no image- or pixellevel\nsegmentation labels. PPBoost first uses vision-language model to produce\ninitial pseudo-bboxes conditioned on the textual object names and applies an\nuncertainty-aware criterion to filter unreliable predictions. The retained imagebboxes\npairs are then leveraged to train a pseudo-labeled detector, producing the\nhigh-quality bboxes for the query images. At inference, PPBoost further refines\nthe generated bboxes by appropriately expand them to tightly cover the target\nanatomical structures. The enhanced spatially-grounding bbox prompts guide existing\nsegmentation models to generate final dense masks, effectively amplifying\nweak text cues into strong spatial guidance. Across three datasets spanning diverse\nmodalities and anatomies, PPBoost consistently improves Dice and Normalized\nSurface Distance over text- and visual-prompted baselines and, notably,\nsurpasses few-shot segmentation models without using labeled data. PPBoost can\ngeneralize to multiple typical visual segmentation model backbones. The anonymous\ncode implementation is in: https://anonymous.4open.science/\nr/submission-code-E2BB/."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Medical Image Segmentation",
"Foundation Model",
"VLM",
"SAM"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/59d2cf254c56cdd90c0fb47f901564179fe3c906.pdf"
},
"primary_area": {
"value": "applications to physical sciences (physics, chemistry, biology, etc.)"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "PPBOOST: PROGRESSIVE PROMPT BOOSTING FOR TEXT-DRIVEN MEDICAL IMAGE SEGMENTATION"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0EXuliYnfW",
"id": "0EXuliYnfW",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission19298/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897047068,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission19298/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission19298/Authors"
]
} | |
2,026 | 0FH7ceYzCq | [
8,
4,
4,
4
] | [
{
"content": "This paper proposes a sequence-agnostic method for continuous multi-modal clustering, Sequence-Agnostic Continual Multi-Modal Clustering (SCMC). It aims to address two core issues in existing continuous multi-modal clustering: the unreliable fusion of historical and new modal information, and the ... | {
"cdate": 1757658463010,
"content": {
"TLDR": {
"value": "We propose a novel sequence-agnostic continual multi-modal clustering method."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025sequenceagnostic,\ntitle={Sequence-agnostic Continual Multi-modal Clustering},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0FH7ceYzCq},\nnote={under review}\n}"
},
"abstract": {
"value": "Continual multi-modal clustering (CMC) aims to address the challenges posed by the continuous arrival of multi-modal data streams, enabling models to progressively update cluster assignments while avoiding catastrophic forgetting.\nCMC closely aligns with the requirements of real-world scenarios and has attracted significant attention from researchers.\nHowever, existing CMC methods face two limitations.\n(1) They fail to reliably model the relationship between historical and new information, leading to redundancy in the shared representation and weakened discriminative power of clustering.\n(2) They are highly sensitive to modality sequence, as early high-quality modalities are gradually forgotten, making the results dependent on the input order.\nTo address these limitations, we propose a novel Sequence-agnostic Continual Multi-modal Clustering (SCMC) method that achieves reliable continual learning and is insensitive to the modality arrival sequence. Specifically, SCMC employs a residual fusion network to suppress the update bias introduced by the newly arrived modalities. It then leverages a cross-temporal knowledge collaboration mechanism to bidirectionally filter information between the historical information and the new modalities, thereby maximizing the preservation of task-relevant information and ensuring reliable continual learning.\nTo eliminate the high sequence sensitivity, we design a sequence-agnostic anti-forgetting strategy, which aligns the current features and cluster distribution with the previous step through cross-temporal consistency transfer, and then prioritizes retaining high-value modality information based on modality importance scores.\nExtensive experiments demonstrate that SCMC outperforms existing SOTA methods, exhibiting sequence insensitivity and strong anti-forgetting capabilities. To the best of our knowledge, SCMC is the first approach to explicitly address the sequence sensitivity problem in CMC."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Multi-modal Clustering",
"Continual Learning"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/6528d0f2a8073d07dfede23027512d9eef7d506e.pdf"
},
"primary_area": {
"value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Sequence-agnostic Continual Multi-modal Clustering"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0FH7ceYzCq",
"id": "0FH7ceYzCq",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission4298/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759898040863,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission4298/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission4298/Authors"
]
} | |
2,026 | 0FJYicpOj0 | [
6,
6,
4,
8
] | [
{
"content": "The paper introduces ε-Gaussian certifiability (GPAR)—a new theoretical notion for analyzing machine unlearning in high-dimensional regimes (p ~ n). It reformulates unlearning guarantees via hypothesis testing and Gaussian trade-off functions, showing that a single Newton step with Gaussian noise ... | {
"cdate": 1758328790343,
"content": {
"TLDR": {
"value": "We introduce the canonical dimension free notion of certifiability suitable to high dimensions and show its utility via a Newton based unlearning algorithm"
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025gaussian,\ntitle={Gaussian certified unlearning in high dimensions: A hypothesis testing approach},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0FJYicpOj0},\nnote={under review}\n}"
},
"abstract": {
"value": "Machine unlearning seeks to efficiently remove the influence of selected data while preserving generalization. Significant progress has been made in low dimensions $(p \\ll n)$, but high dimensions pose serious theoretical challenges as standard optimization assumptions of $\\Omega(1)$ strong convexity and $O(1)$ smoothness of the per-example loss $f$ rarely hold simultaneously in proportional regimes $(p\\sim n)$.\nIn this work, we introduce $\\varepsilon$-Gaussian certifiability, a canonical and robust notion well-suited to high-dimensional regimes, that optimally captures a broad class of noise adding mechanisms. Then we theoretically analyze the performance of a widely used unlearning algorithm based on one step of the Newton method in the high-dimensional setting described above. Our analysis shows that a single Newton step, followed by a well-calibrated Gaussian noise, is sufficient to achieve both privacy and accuracy in this setting. This result stands in sharp contrast to the only prior work that analyzes machine unlearning in high dimensions \\citet{zou2025certified}, which relaxes some of the standard optimization assumptions for high-dimensional applicability, but operates under the notion of $\\varepsilon$-certifiability. That work concludes %that a single Newton step is insufficient even for removing a single data point, and\nthat at least two steps are required to ensure both privacy and accuracy. Our result leads us to conclude that the discrepancy in the number of steps arises because of the sub optimality of the notion of $\\varepsilon$-certifiability and its incompatibility with noise adding mechanisms, which $\\varepsilon$-Gaussian certifiability is able to overcome optimally."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Machine unlearning in high dimensions",
"Proportional asymptotics",
"High dimensional statistical theory",
"Privacy–accuracy tradeoff",
"Hypothesis testing",
"Gaussian noise calibration",
"Newton method"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/2aa1ed1036190b875838c897c261f703465c440f.pdf"
},
"primary_area": {
"value": "alignment, fairness, safety, privacy, and societal considerations"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Gaussian certified unlearning in high dimensions: A hypothesis testing approach"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0FJYicpOj0",
"id": "0FJYicpOj0",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission22272/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759896875746,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission22272/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission22272/Authors"
]
} | |
2,026 | 0FN0u6qTAi | [
4,
4,
2,
6
] | [
{
"content": "This paper introduces the \"Protein-as-Second-Language\" framework, which aims to enable large language models (LLMs) to interpret protein (amino acid) sequences as if they were acquiring a second symbolic language. By curating a bilingual dataset of almost 80k protein-question-answer triples and ... | {
"cdate": 1757862318304,
"content": {
"TLDR": {
"value": "We propose a protein–language framework and bilingual dataset that enable LLMs to reason about protein function via context-driven learning without fine-tuning."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025protein,\ntitle={Protein as a Second Language for {LLM}s},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0FN0u6qTAi},\nnote={under review}\n}"
},
"abstract": {
"value": "Deciphering the function of unseen protein sequences is a fundamental challenge with broad scientific impact, yet most existing methods depend on task-specific adapters or large-scale supervised fine-tuning. We introduce the \"Protein-as-Second-Language\" framework, which reformulates amino-acid sequences as sentences in a novel symbolic language that large language models can interpret through contextual exemplars. Our approach adaptively constructs sequence–question–answer triples that reveal functional cues in a zero-shot setting, without any further training. To support this process we curate a bilingual corpus of 79,926 protein–QA instances spanning attribute prediction, descriptive understanding, and extended reasoning. Empirically, our method delivers consistent gains across diverse open-source LLMs and GPT-4o, achieving up to 17.2% ROUGE-L improvement (average +7%) and even surpassing fine-tuned protein-specific language models. These results highlight that generic LLMs, when guided with protein-as-language cues, can outperform domain-specialized models, offering a scalable pathway for protein understanding in foundation models."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Large language models; Protein–QA dataset; Context-Driven Learning; Zero-shot learning"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/68dc93b54dfdeefd77080a635770b44d29c1e969.pdf"
},
"primary_area": {
"value": "datasets and benchmarks"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Protein as a Second Language for LLMs"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0FN0u6qTAi",
"id": "0FN0u6qTAi",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission5183/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897990156,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission5183/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission5183/Authors"
]
} | |
2,026 | 0Fc9yLlIYX | [
4,
2,
2,
4
] | [
{
"content": "This paper proposes a systematic and comprehensive analysis of temporal bias in Large Audio Language Models (LALMs). Through experiments, the paper reveals that LALMs consistently predict the temporal occurrence of acoustic events earlier. Through detailed analysis, the paper shows that in LALMs: ... | {
"cdate": 1758297559979,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025lost,\ntitle={Lost in Time: Systematic Temporal Bias in Large Audio Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0Fc9yLlIYX},\nnote={under review}\n}"
},
"abstract": {
"value": "Large Audio Language Models (LALMs) are increasingly applied to audio understanding and multimodal reasoning, yet their ability to locate when events occur remains underexplored. We present the first systematic study of temporal bias in LALMs, revealing a key limitation in their timestamp prediction. For example, when asked “At which second does the lecturer introduce the key formula?”, models often predict timestamps that are consistently earlier or later than the ground truth. Through controlled experiments on timestamped datasets, we find that temporal bias (i) is prevalent across datasets and models, (ii) increases with audio length—even accumulating to tens of seconds in extended recordings, and (iii) varies across event types and positions. We quantify this effect with the Temporal Bias Index (TBI), measuring systematic misalignment in predicted event timings, and complement it with a visualization framework. Our findings highlight a fundamental limitation in current LALMs and call for the development of temporally robust architectures."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Large Audio Language Models",
"Temporal Bias",
"Audio Understanding"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/e383f2c4c0f2a00cde1c8b0496044e619c38268b.pdf"
},
"primary_area": {
"value": "interpretability and explainable AI"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Lost in Time: Systematic Temporal Bias in Large Audio Language Models"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0Fc9yLlIYX",
"id": "0Fc9yLlIYX",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission19599/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897031002,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission19599/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission19599/Authors"
]
} | |
2,026 | 0FhrtdKLtD | [
6,
6,
6,
2
] | [
{
"content": "This paper introduces MindCube, a new benchmark for evaluating the ability of VLMs to form \"spatial mental models\" from limited views. The authors show that existing models perform poorly, then propose a \"map-then-reason\" approach, which is to train a model to first generate a structured cogni... | {
"cdate": 1757013520622,
"content": {
"TLDR": {
"value": "We propose MindCube and find existing VLMs perform poorly on it. Supervising models to first generate cognitive maps and then reason upon them proves to be a quite effective approximation for spatial mental modeling from limited views."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025understanding,\ntitle={Understanding {VLM}s Spatial Mental Modeling Capability from Limited Views},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0FhrtdKLtD},\nnote={under review}\n}"
},
"abstract": {
"value": "Can Vision Language Models (VLMs) imagine the full scene from just a few views, like humans do? Humans form spatial mental models, internal representations of unseen space, to reason about layout, perspective, and motion. Our new MindCube benchmark with 21,154 questions across 3,268 images exposes this critical gap, where existing VLMs exhibit near-random performance. Using MindCube, we systematically evaluate how well VLMs build robust spatial mental models through representing positions (cognitive mapping), orientations (perspective-taking), and dynamics (mental simulation for \"what-if\" movements). We then explore three approaches to help VLMs approximate spatial mental models, including unseen intermediate views, natural language reasoning chains, and cognitive maps. The significant improvement comes from a synergistic approach, \"map-then-reason\", that jointly trains the model to first generate a cognitive map and then reason upon it. By training models to reason over these internal maps, we boosted accuracy from 37.8% to 60.8% (+23.0%). Adding reinforcement learning pushed performance even further to 70.7% (+32.9%). Our key insight is that such scaffolding of spatial mental models, actively constructing and utilizing internal structured spatial representations with flexible reasoning processes, significantly improves understanding of unobservable space."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Vision Language Models",
"VLMs",
"Multi Modal Language Models",
"Spatial Intelligence",
"Spatial Reasoning"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/b3ae14a291ec47f47838c66b9d91f330cab8c231.pdf"
},
"primary_area": {
"value": "foundation or frontier models, including LLMs"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/5082d91d065fb9b9d69c480d6f04be96d47a8858.zip"
},
"title": {
"value": "Understanding VLMs Spatial Mental Modeling Capability from Limited Views"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0FhrtdKLtD",
"id": "0FhrtdKLtD",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission2183/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759898164442,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission2183/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission2183/Authors"
]
} | |
2,026 | 0G8Cq9z2Hp | [
4,
4,
4,
6,
6
] | [
{
"content": "This paper addresses the computational complexity issue in AlphaFold by presenting a hierarchical pipeline, refered to as HieraFold, which decomposes the end-to-end structure prediction task in a coarse-to-fine manner.\nHieraFold first performs a coarse global prediction using a \"lightweight\" ve... | {
"cdate": 1758359066119,
"content": {
"TLDR": {
"value": "We introduce HierAFold, a hierarchical pipeline that exploits the modularity of large complexes via PAE-guided (Predicted Aligned Error) subunit decomposition, targeted interface-aware refinement, and confidence-weighted assembly."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025efficient,\ntitle={Efficient Prediction of Large Protein Complexes via Subunit-Guided Hierarchical Refinement},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0G8Cq9z2Hp},\nnote={under review}\n}"
},
"abstract": {
"value": "State-of-the-art protein structure predictors have revolutionized structural biology, yet quadratic memory growth with token length makes end-to-end inference impractical for large complexes beyond a few thousand tokens. We introduce \\textsc{HierAFold}, a hierarchical pipeline that exploits the modularity of large complexes via PAE-guided (Predicted Aligned Error) subunit decomposition, targeted interface-aware refinement, and confidence-weighted assembly. PAE maps localize rigid intra-chain segments and sparse inter-chain interfaces, enabling joint refinement of likely interacting subunits to capture multi-body cooperativity without increasing memory. \\textsc{HierAFold} matches AlphaFold3 accuracy, raises success rates from 49.9\\% (CombFold) to 73.1\\% on recent PDB set. While for large complexes, it cuts peak memory by $\\sim$25\\,GB on a 4{,}000-token target ($\\sim$40\\%), successfully models complexes with over $5{,}000$ tokens that are out-of-memory for AlphaFold3, and raises success rates by two-fold compared with CombFold."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Protein complex structure prediction",
"AlphaFold3",
"complex modularity"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/cb9c051ccacad225ae82bee4a700f0e28025501a.pdf"
},
"primary_area": {
"value": "applications to physical sciences (physics, chemistry, biology, etc.)"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Efficient Prediction of Large Protein Complexes via Subunit-Guided Hierarchical Refinement"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0G8Cq9z2Hp",
"id": "0G8Cq9z2Hp",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission24657/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759896756603,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission24657/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission24657/Authors"
]
} | |
2,026 | 0GMt2OWeCb | [
4,
4,
2,
2
] | [
{
"content": "This paper addresses two critical limitations of existing memory-augmented Large Language Model (LLM)-based agents: low data efficiency (relying on extensive task-specific interaction data for early training) and poor adaptability (using static memory retrieval strategies that fail to balance cros... | {
"cdate": 1758343379022,
"content": {
"TLDR": {
"value": "We propose a memory-augmented LLM agent with cross-task learning and dynamic memory retrieval to improve adaptability and efficiency in multi-turn instruction-following tasks."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025memoryaugmented,\ntitle={Memory-Augmented Large Language Model-Based Agent with Cross-Task Experience Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0GMt2OWeCb},\nnote={under review}\n}"
},
"abstract": {
"value": "Large Language Model (LLM)-based agents have demonstrated impressive capabilities in complex decision-making and multi-turn instruction-following tasks. To enhance knowledge retention and contextual adaptability, recent work has equipped these agents with memory modules that store and reuse historical interaction experiences. However, existing memory-augmented approaches face two key limitations: they often require large amounts of interaction data during early training to reach competitive performance, resulting in low data efficiency; and they rely on static, self-derived experience reuse strategies, limiting their ability to adapt when prior learning is insufficient and preventing the use of transferable knowledge from related tasks. Building on these observations, in this paper, we propose a memory-augmented LLM agent with cross-task experience learning, designed to improve data efficiency and adaptability. Our method augments the conventional task-specific memory with an additional source experience memory that retains transferable knowledge from related but distinct tasks. We further introduce a dynamic memory retrieval mechanism that adaptively draws from both task and source memories, allowing the agent to balance prior task-specific experiences with cross-task knowledge according to the current context and progression. We validate the proposed method on the WebShop benchmark, which comprises diverse, multi-turn instruction-following tasks across product domains with varying semantic complexity. Experimental results show that our approach consistently outperforms state-of-the-art memory-augmented LLM agents in task success rate and generalization, demonstrating the effectiveness of the proposed memory architecture and retrieval mechanism."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Large Language Models",
"LLM-based Agents",
"Experience Transfer",
"Long-Term Memory Mechanisms"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/a643457b43dfbecfbe3e4273007e64a36d4b82c4.pdf"
},
"primary_area": {
"value": "transfer learning, meta learning, and lifelong learning"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Memory-Augmented Large Language Model-Based Agent with Cross-Task Experience Learning"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0GMt2OWeCb",
"id": "0GMt2OWeCb",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission23410/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759896816282,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission23410/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission23410/Authors"
]
} | |
2,026 | 0GNBqoYcAP | [
4,
4,
6
] | [
{
"content": "his paper examines how world models can learn and adapt through context, focusing on in-context learning (ICL) within both MDP and POMDP settings. The authors distinguish between two key processes, namely In-Context Environment Learning (ICEL) and In-Context Environment Recognition (ICER). Further... | {
"cdate": 1758188335809,
"content": {
"TLDR": {
"value": "We formalize, bound, and validate in-context environment learning, showing that long-context, diverse-input world models can self-adapt by recognizing or learning new dynamics without parameter updates."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025context,\ntitle={Context and Diversity Matter: The Emergence of In-Context Learning in World Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0GNBqoYcAP},\nnote={under review}\n}"
},
"abstract": {
"value": "The capability of predicting environmental dynamics underpins both biological neural systems and general embodied AI in adapting to their surroundings. Yet prevailing approaches rest on static world models that falter when confronted with novel or rare configurations. We investigate in-context environment learning (ICEL), shifting attention from zero-shot performance to the growth and asymptotic limits of the world model. Our contributions are three-fold: (1) we formalize in-context learning of a world model and identify two core mechanisms: environment recognition and environment learning; (2) we derive error upper-bounds for both mechanisms that expose how the mechanisms emerge; and (3) we empirically confirm that distinct ICL mechanisms exist in the world model, and we further investigate how data distribution and model architecture affect ICL in a manner consistent with theory. These findings demonstrate the potential of self-adapting world models and highlight the key factors behind the emergence of ICEL, most notably the necessity of long context and diverse environments."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"In-Context Learning; World Models"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/f6b6f2acb5612c1bcb18dd2a927c42e5e641931b.pdf"
},
"primary_area": {
"value": "transfer learning, meta learning, and lifelong learning"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/5003af42abf06668a2f7d72aec3b310657d4c3cc.zip"
},
"title": {
"value": "Context and Diversity Matter: The Emergence of In-Context Learning in World Models"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0GNBqoYcAP",
"id": "0GNBqoYcAP",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission11055/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897611882,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission11055/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission11055/Authors"
]
} | |
2,026 | 0GaCfBRFnf | [
6,
4,
6,
8
] | [
{
"content": "This paper introduces ProActive Self-Refinement (PASR) as a novel method for enabling Large Language Models (LLMs) to refine their outputs during the generation process, rather than as a post-hoc step. \n\nThe authors formalize this as a MDP and use RL to train models to decide whether, when, and ... | {
"cdate": 1758282395015,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025a,\ntitle={A Stitch in Time Saves Nine: Proactive Self-Refinement for Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0GaCfBRFnf},\nnote={under review}\n}"
},
"abstract": {
"value": "Recent advances in self-refinement have demonstrated significant potential for improving the outputs of large language models (LLMs) through iterative refinement. However, most existing self-refinement methods rely on a reactive process with a fixed number of iterations, making it difficult to determine the optimal timing and content of refinement based on the evolving generation context. Inspired by the way humans dynamically refine their thoughts during execution, we propose ProActive Self-Refinement (PASR), a novel method that enables LLMs to refine their outputs during the generation process. Unlike methods that regenerate entire responses, PASR proactively decides whether, when, and how to refine based on the model’s internal state and evolving context. We conduct extensive experiments on a diverse set of 10 tasks to evaluate the effectiveness of PASR. Experimental results show that PASR significantly enhances problem-solving performance. In particular, on Qwen3-8B, PASR reduces average token consumption by 41.6% compared to standard generation, while also achieving an 8.2% improvement in accuracy. Our code and all baselines used in the paper are available in the GitHub."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Large language models",
"Self-refine"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/bc704b7bedab5c7f854cc8447788460efbeba2e4.pdf"
},
"primary_area": {
"value": "applications to computer vision, audio, language, and other modalities"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "A Stitch in Time Saves Nine: Proactive Self-Refinement for Language Models"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0GaCfBRFnf",
"id": "0GaCfBRFnf",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission17956/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897142928,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission17956/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission17956/Authors"
]
} | |
2,026 | 0GdjEJCHOE | [
6,
2,
2,
2
] | [
{
"content": "This paper presents DRMLP, a Dynamic Regularized Multi-Layer Perceptron framework for discovering Granger causal structure in multivariate time series. DRMLP introduces a dual-branch neural architecture, combining a linear (MLP-based) causal discovery path with a recurrent (LSTM-based) sampling st... | {
"cdate": 1757163600341,
"content": {
"TLDR": {
"value": "A dynamic regularization approach for Granger-based causal discovery achieves superior performance on simulated and real-world time series data."
},
"_bibtex": {
"value": "@misc{\nliu2025drmlp,\ntitle={{DRMLP}: Dynamic Regularized Multi-Layer Perceptron for Neural Granger Causality Discovery with Adaptive Temporal Penalties},\nauthor={Haiyang Liu and Wenrui Jiang and Xiaokang Wang and Muyun Yao},\nyear={2025},\nurl={https://openreview.net/forum?id=0GdjEJCHOE}\n}"
},
"abstract": {
"value": "With the rapid development of IoT devices, collecting multivariate time series data has become increasingly convenient. Understanding the causal relationships among different time series variables is critical for validating causal discovery methods and benchmarking their ability to recover ground-truth interactions in controlled synthetic environments. However, existing Granger causality approaches based on neural networks typically require modeling each time series variable separately and assume that the influence of historical values decays over time. These limitations result in complex models and poor performance in discovering causality in time series with long-range dependencies. To address these drawbacks, this paper proposes a model called DRMLP: Dynamic Regularized Multi-Layer Perceptron, a Granger causality discovery method capturing periodic temporal dependencies from the input weights of a convolutional network. The proposed approach employs a dual-branch neural network architecture: a linear causal discovery network is utilized to extract causal relations from sampled weight data, while a hierarchical regularization strategy is introduced to optimize the weights of the convolutional network. This design enhances the accuracy of causal relation discovery and reduces noise interference, thereby ensuring the temporal consistency of the identified causal structures. Experiments conducted on simulated datasets and real-world system-generated datasets show that DRMLP outperforms state-of-the-art baseline methods."
},
"anonymous_url": null,
"authorids": {
"value": [
"~Haiyang_Liu3",
"~Wenrui_Jiang1",
"~Xiaokang_Wang6",
"~Muyun_Yao1"
]
},
"authors": {
"value": [
"Haiyang Liu",
"Wenrui Jiang",
"Xiaokang Wang",
"Muyun Yao"
]
},
"code_of_ethics": null,
"keywords": {
"value": [
"time series",
"causal discovery",
"deep learning",
"regularization",
"mlp"
]
},
"no_acknowledgement_section": null,
"paperhash": {
"value": "liu|drmlp_dynamic_regularized_multilayer_perceptron_for_neural_granger_causality_discovery_with_adaptive_temporal_penalties"
},
"pdf": {
"value": "/pdf/8da9719f16b180456bce3c0635cf484d9aaaadd6.pdf"
},
"primary_area": {
"value": "causal reasoning"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "DRMLP: Dynamic Regularized Multi-Layer Perceptron for Neural Granger Causality Discovery with Adaptive Temporal Penalties"
},
"venue": {
"value": "ICLR 2026 Conference Withdrawn Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Withdrawn_Submission"
}
},
"forum": "0GdjEJCHOE",
"id": "0GdjEJCHOE",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/-/Withdrawn_Submission"
],
"license": "CC BY 4.0",
"mdate": 1763026722172,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission2610/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission2610/Authors"
]
} | |
2,026 | 0GjORP5Duq | [
4,
6,
2,
6
] | [
{
"content": "The paper addresses the persistent challenge of compositional reasoning in vision–language models such as CLIP. \nIt proposes RACA-CLIP, a structured contrastive learning framework that integrates scene-graph supervision to align visual and textual representations at the object, attribute, and rel... | {
"cdate": 1758352745950,
"content": {
"TLDR": {
"value": "Building compositionality robust CLIP model by region aware training objectives, pushing them towards better reasoning."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025racaclip,\ntitle={{RACA}-{CLIP}: Relation-Aware Compositional Alignment for {CLIP}},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0GjORP5Duq},\nnote={under review}\n}"
},
"abstract": {
"value": "Vision-Language Models (VLMs) such as CLIP excel at broad multimodal tasks, yet struggle with compositional reasoning. Despite capturing coarse correlations, they often act like “bags-of-words” missing fine-grained structures such as object–attribute bindings and inter-object relations. We attribute this to: (i) limited compositional diversity in large-scale image–text data, and (ii) contrastive objectives that emphasize global alignment over grounded structure. To address this, we propose a hierarchical fine-grained alignment framework that explicitly bridges visual and textual components at the object, attribute, and relation levels. Unlike prior work relying on parsers, we leverage scene graph annotated datasets for structured supervision, requiring no extra labeling. We introduce a hierarchical fine-grained loss to complement standard contrastive learning by grounding entities and relations across modalities. Experiments on compositional benchmarks SugarCrepe, What’sUp, and Cola show large gains in capturing nuanced structure, while preserving performance on standard vision-language tasks. RACA CLIP method improves compositional reasoning accuracy by +24.86% on SugarCrepe, +5.7% on What’sUp, and +4.76 on Cola, offering a simple yet effective path toward stronger, human-like compositional understanding."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Explainablity",
"Vision-Language Models",
"Compositionality"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/9bd2b282be01ace8b9e99b998f90c76c516b8eed.pdf"
},
"primary_area": {
"value": "interpretability and explainable AI"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "RACA-CLIP: Relation-Aware Compositional Alignment for CLIP"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0GjORP5Duq",
"id": "0GjORP5Duq",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission24104/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759896781548,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission24104/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission24104/Authors"
]
} | |
2,026 | 0GlStRq4Xw | [
0,
8,
2,
6
] | [
{
"content": "This paper proposes a machine learning architecture for constrained optimization learning that approximates an iterative descent algorithm. The proposed approach integrates an active set strategy, an approximate descent direction computation, and a projection operator to ensure equality constraint... | {
"cdate": 1757218565386,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025descentnet,\ntitle={Descent-Net: Learning Descent Directions for Constrained Optimization},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0GlStRq4Xw},\nnote={under review}\n}"
},
"abstract": {
"value": "Deep learning approaches, known for their ability to model complex relationships and fast execution, are increasingly being applied to solve large optimization problems. However, existing methods often face challenges in simultaneously ensuring feasibility and achieving an optimal objective value. To address this issue, we propose Descent-Net, a neural network designed to learn an effective descent direction from a feasible solution. By updating the solution along this learned direction, Descent-Net improves the objective value while preserving feasibility. Our method demonstrates strong performance on both synthetic optimization tasks and the real-world AC optimal power flow problem."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"descent direction",
"unrolling",
"L2O",
"constrained optimization"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/9f80b2167f277c1913f7e3983be25ea3c1ecb2e5.pdf"
},
"primary_area": {
"value": "optimization"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/6fdadc0816e37f5892c278ac3d9df6c5837c9786.zip"
},
"title": {
"value": "Descent-Net: Learning Descent Directions for Constrained Optimization"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0GlStRq4Xw",
"id": "0GlStRq4Xw",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission2713/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759898132028,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission2713/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission2713/Authors"
]
} | |
2,026 | 0GpolO2auw | [
8,
6,
4
] | [
{
"content": "This paper considers the task of community detection in well-clusterable graphs with sublinear space. The goal is to design a data structure D that fits in sublinear memory, and that enables one to query the cluster assignment for each node in sublinear time. Previous approaches all require $\\Ome... | {
"cdate": 1758270235850,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025sublinear,\ntitle={Sublinear Spectral Clustering Oracle with Little Memory},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0GpolO2auw},\nnote={under review}\n}"
},
"abstract": {
"value": "We study the problem of designing *sublinear spectral clustering oracles* for well-clusterable graphs. Such an oracle is an algorithm that, given query access to the adjacency list of a graph $G$, first constructs a compact data structure $\\mathcal{D}$ that captures the clustering structure of $G$. Once built, $\\mathcal{D}$ enables sublinear time responses to \\textsc{WhichCluster}$(G,x)$ queries for any vertex $x$. A major limitation of existing oracles is that constructing $\\mathcal{D}$ requires $\\Omega(\\sqrt{n})$ memory, which becomes a bottleneck for massive graphs and memory-limited settings. In this paper, we break this barrier and establish a memory-time trade-off for sublinear spectral clustering oracles. Specifically, for well-clusterable graphs, we present oracles that construct $\\mathcal{D}$ using much smaller than $O(\\sqrt{n})$ memory (e.g., $O(n^{0.01})$) while still answering membership queries in sublinear time. We also characterize the trade-off frontier between memory usage $S$ and query time $T$, showing, for example, that $S\\cdot T=\\widetilde{O}(n)$ for clusterable graphs with a logarithmic conductance gap, and we show that this trade-off is nearly optimal (up to logarithmic factors) for a natural class of approaches. Finally, to complement our theory, we validate the performance of our oracles through experiments on synthetic networks."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Graph Clustering",
"Spectral Clustering",
"Memory-Efficient Algorithms",
"Sublinear Algorithms",
"Space-Time Trade-offs"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/16c54ec8b3c8fd02e67e16d05874f80566e17f23.pdf"
},
"primary_area": {
"value": "learning theory"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Sublinear Spectral Clustering Oracle with Little Memory"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0GpolO2auw",
"id": "0GpolO2auw",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission16919/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897210152,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission16919/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission16919/Authors"
]
} | |
2,026 | 0H5iD4he7R | [
2,
6,
6,
2
] | [
{
"content": "The paper proposes f-DMU, a unified framework for diffusion model unlearning based on f-divergence. It generalizes existing MSE-based and KL-based unlearning approaches by allowing any f-divergence. The method provides two formulations—closed-form and variational—to balance simplicity and generali... | {
"cdate": 1758357192051,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025a,\ntitle={A Unified Framework for Diffusion Model Unlearning with f-Divergence},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0H5iD4he7R},\nnote={under review}\n}"
},
"abstract": {
"value": "Machine unlearning aims to remove specific knowledge from a trained model. While diffusion models (DMs) have shown remarkable generative capabilities, existing unlearning methods for text-to-image (T2I) models often rely on minimizing the mean squared error (MSE) between the output distribution of a target and an anchor concept. \nWe show that this MSE-based approach is a special case of a unified $f$-divergence-based framework, in which any $f$-divergence can be utilized.\nWe analyze the benefits of using different $f$-divergences, that mainly impact the convergence properties of the algorithm and the quality of unlearning. \nThe proposed unified framework offers a flexible paradigm that allows to select the optimal divergence for a specific application, balancing different trade-offs between aggressive unlearning and concept preservation."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"machine unlearning",
"diffusion models",
"f-divergence"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/fdc476126ec69532b1dc4e0377d2bc0e0aa18520.pdf"
},
"primary_area": {
"value": "alignment, fairness, safety, privacy, and societal considerations"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/d71c45a22db08d53a7c3b022703ec331a121c4c8.zip"
},
"title": {
"value": "A Unified Framework for Diffusion Model Unlearning with f-Divergence"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0H5iD4he7R",
"id": "0H5iD4he7R",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission24471/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759896764314,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission24471/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission24471/Authors"
]
} | |
2,026 | 0HcqZkv1zs | [
4,
4,
4,
8
] | [
{
"content": "The authors propose a method to incorporate semantic context into theorem proving models. They propose a clarity score to evaluate the understanding of this context. They demonstrate that this clarity helps downstream performance in proving theorems.",
"id": "XnDxg2qrMS",
"rating": 4
},
... | {
"cdate": 1758161650677,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025clarifying,\ntitle={Clarifying Before Reasoning: A Coq Prover with Structural Context},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0HcqZkv1zs},\nnote={under review}\n}"
},
"abstract": {
"value": "In this work, we investigate whether improving task clarity can enhance reasoning ability of large language models, focusing on theorem proving in Coq. We introduce a concept-level metric to evaluate task clarity and show that adding structured semantic context to the standard input used by modern LLMs, leads to a 1.85$\\times$ improvement in clarity score (44.5\\%~$\\rightarrow$~82.3\\%). Using the general-purpose model DeepSeek-V3, our approach leads to a 2.1$\\times$ improvement in proof success (21.8\\%~$\\rightarrow$~45.8\\%) and outperforms the previous state-of-the-art Graph2Tac (33.2\\%). We evaluate this on 1,386 theorems randomly sampled from 15 standard Coq packages, following the same evaluation protocol as Graph2Tac.\nFurthermore, fine-tuning smaller models on our structured data can achieve even higher performance (48.6\\%).\nOur method uses selective concept unfolding to enrich task descriptions, and employs a Planner-Executor architecture. These findings highlight the value of structured task representations in bridging the gap between understanding and reasoning."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"theorem proving",
"Coq",
"structured reasoning",
"formal verification"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/ead97ffe8496c0912157e9ef7e4234c8823ec935.pdf"
},
"primary_area": {
"value": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Clarifying Before Reasoning: A Coq Prover with Structural Context"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0HcqZkv1zs",
"id": "0HcqZkv1zs",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission10135/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897671780,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission10135/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission10135/Authors"
]
} | |
2,026 | 0I2N8KxOAo | [
2,
2,
2,
6
] | [
{
"content": "This paper proposes DeFa, a framework for non-stationary multivariate time series forecasting. DeFa consists of two main components: (1) NAILong, a decomposition strategy that separates a time series into a time-varying Amplifier, normalized Seasonality, and sparse Residual via a multiplicative fo... | {
"cdate": 1756731475441,
"content": {
"TLDR": {
"value": "DeFA introduces a decomposition-based framework with tensor autoregressive forecasting that effectively captures non-stationary dynamics and long-term dependencies in multivariate time series."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025defa,\ntitle={DeFa: Non-Stationary Decomposition and Factorized Forecasting for Multivariate Time Series},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0I2N8KxOAo},\nnote={under review}\n}"
},
"abstract": {
"value": "Multivariate time series forecasting is essential in fields like energy systems, weather prediction, and traffic monitoring. While recent deep learning models, including Transformer-based architectures, show potential, they often struggle to capture the complex dynamics and non-stationary patterns inherent in real-world data. This limitation arises from over-parametrization and the difficulty in modelling shifting patterns in simple short- and long-term terms. In this paper, we propose a unified framework, DeFa, that addresses these challenges by combining decomposition-based modelling with tensor autoregressive forecasting. To capture long-term dynamics, stationary seasonality, and sparse residuals unique to non-stationary time series, DeFa decomposes the input series into three components using the Non-stationary AdaptiveInteractive Long-term strategy (NAILong). Furthermore, to improve the prediction of the Amplifier, which encodes time-varying dynamics, DeFa is enhanced with the Factorized Tensor Autoregression framework (FaTA). Unlike existing methods that disentangle or represent input series directly, FaTA explicitly models the autoregressive coefficient tensor across variates and temporal dimensions. This fusion enables a more flexible and interpretable representation of multi-variable interactions, improving forecasting accuracy while maintaining computational efficiency. Extensive experiments on real-world datasets show that DeFa outperforms state-of-the-art methods in terms of both interpretable forecasting accuracy and scalability. Additionally, DeFa handles long-term dynamics and drifting seasonalities efficiently through a plug-in option, extending its adaptability."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"time series forecasting",
"deep learning"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/b94cded0e13a77d167ddff7dbbf4055ceb64596c.pdf"
},
"primary_area": {
"value": "learning on time series and dynamical systems"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "DeFa: Non-Stationary Decomposition and Factorized Forecasting for Multivariate Time Series"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0I2N8KxOAo",
"id": "0I2N8KxOAo",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission217/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759898271392,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission217/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission217/Authors"
]
} | |
2,026 | 0IFqBfX7Ak | [
4,
2,
6
] | [
{
"content": "This paper introduces Integrated Policy Gradient (IPG), a method intended to attribute and modulate reasoning components in large language models by applying a policy‐gradient–like formulation on hidden activations, followed by scaling of the identified components. The aim is to locate “reasoning ... | {
"cdate": 1758256779840,
"content": {
"TLDR": {
"value": "A method to causally control and interpret LLM reasoning behaviors by identifying and intevening internal reasoning-critical components."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025interpreting,\ntitle={Interpreting and Controlling {LLM} Reasoning through Integrated Policy Gradient},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0IFqBfX7Ak},\nnote={under review}\n}"
},
"abstract": {
"value": "Large language models (LLMs) demonstrate strong reasoning abilities in solving complex real-world problems. Yet, the internal mechanisms that support these behaviors remain opaque, raising concerns regarding truthfulness, safety, and controllability in practical applications.\nExisting interpretability approaches either rely on human-annotated contrastive pairs to derive control vectors, which limits reliability and generalization, or identify neurons correlated with superficial textual concepts, failing to capture the complexity of reasoning processes. \nConsequently, current methods struggle to precisely localize complex reasoning mechanisms or capture causal effects from model internal workings to the reasoning outputs.\nIn this paper, we build on causality-aware and outcome-oriented principles that focus on identifying components that have causal contributions to reasoning behavior where outcomes are cumulated by long-range effects.\nWe propose Integrated Policy Gradient (IPG), a novel framework that attributes reasoning behaviors to model inner workings like neurons, by propagating compound outcome-based signals (e.g., post reasoning accuracy) backward through model inference trajectories.\nIPG is efficient requiring only a few calls to the standard gradient operator, which uncovers causal structures governing complex reasoning and avoids large manual supervision.\nEmpirical evaluations demonstrate that our approach achieves more precise mechanistic interpretability and enables reliable modulation of reasoning behaviors across diverse reasoning models."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Large Language Models",
"Reasoning",
"Mechanistic Interpretability",
"Policy Gradient",
"Sparse Autoencoder"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/37cc955be40917d01dca561f285b20133300783f.pdf"
},
"primary_area": {
"value": "interpretability and explainable AI"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Interpreting and Controlling LLM Reasoning through Integrated Policy Gradient"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0IFqBfX7Ak",
"id": "0IFqBfX7Ak",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission15897/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897274656,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission15897/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission15897/Authors"
]
} | |
2,026 | 0IN8RiFbmg | [
4,
4,
2,
2
] | [
{
"content": "This paper investigates the performance of Parameter-Efficient Fine-Tuning (PEFT) methods under increasing distribution shifts across tasks. We introduce a novel PEFT technique, AUG, which augments matrix-vector products with learnable parameters conditioned on both the input data and pretrained w... | {
"cdate": 1758158798611,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025scaling,\ntitle={Scaling Parameter-Efficiency with Distribution Shifts for Domain Adaptation},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0IN8RiFbmg},\nnote={under review}\n}"
},
"abstract": {
"value": "Distribution shifts between source and target domains pose significant challenges to the generalization capabilities of machine learning models. While foundation models are often fine-tuned to adapt to new domains, their increasing size has led to a rise in the computational resources required for domain adaptation. This has driven interest in Parameter-Efficient Fine-Tuning (PEFT) methods, which have shown strong performance on in-domain tasks. In this work, we investigate how PEFT methods scale with varying degrees of distribution shifts and propose a novel PEFT method designed for domain adaptation. We select an English pre-trained Large Language Model (LLM) as the foundation model and apply PEFT techniques across tasks that progressively introduce larger distribution shifts. Specifically, we begin with SuperGLUE English benchmark, followed by a multilingual inference task for high-resource and low-resource languages, then a multimodal image captioning task. Finally, We introduce a novel multimodal and multitemporal radar interferometry task for detecting charcoal production sites in remote areas. Separately, we propose a PEFT method that augments matrix vector products with learnable parameters, inducing a learning paradigm that conditions on both training data and encoded information. Our method is competitive against SOTA PEFT methods for English tasks and out-performs SOTA methods for larger distribution shifts i.e. low-resource multilingual, image captioning, and radar interferometry tasks."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"peft",
"remote-sensing"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/26c0d3120b9a28d108df43635c02ee2e81499105.pdf"
},
"primary_area": {
"value": "applications to computer vision, audio, language, and other modalities"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Scaling Parameter-Efficiency with Distribution Shifts for Domain Adaptation"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0IN8RiFbmg",
"id": "0IN8RiFbmg",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission10053/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897678056,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission10053/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission10053/Authors"
]
} | |
2,026 | 0IWZjbMmry | [
4,
4,
2,
2
] | [
{
"content": "This paper introduces LayerDecompose, a compression framework for large language models that combines weight sharing with low-rank adapters. The key idea is to represent groups of consecutive layers with a single shared weight matrix W, augmented with layer-specific low-rank residuals and per-chan... | {
"cdate": 1758267368176,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025layerdecompose,\ntitle={LayerDecompose: Exploring weight sharing for Large Language Model Compression},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0IWZjbMmry},\nnote={under review}\n}"
},
"abstract": {
"value": "Recent advances in large language model (LLM) compression have predominantly focused on pruning and low-rank factorization, leaving weight sharing—despite its success in classical neural network compression—largely unexplored. We introduce LayerDecompose, a novel framework that reduces parameter redundancy by sharing a core weight matrix across transformer layers and augmenting each layer with lightweight, low-rank adapters. Unlike prior SVD- and pruning-based methods, our joint optimization of shared weights and residual adapters achieves a 30% model size reduction while retaining 89% of the original performance on seven standard benchmarks. Experiments on LLaMA and other models demonstrate that LayerDecompose consistently outperforms state-of-the-art baselines. These results highlight the promise of combining weight sharing with low-rank adaptation for efficient, scalable LLM deployment."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Large Language Models (LLMs)",
"Model Compression",
"Weight Sharing"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/8cec12cc44b2a6394692a3239cbd59726b460b02.pdf"
},
"primary_area": {
"value": "foundation or frontier models, including LLMs"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "LayerDecompose: Exploring weight sharing for Large Language Model Compression"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0IWZjbMmry",
"id": "0IWZjbMmry",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission16659/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897226556,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission16659/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission16659/Authors"
]
} | |
2,026 | 0IceiDrfxI | [
4,
2,
4,
4
] | [
{
"content": "This paper introduces NATLM, a framework that combines static analysis (AST/CFG) and LLM reasoning (Gemini Pro 1.5) to detect four NFT smart-contract defect types: ERC-721 Reentrancy, Public Burn, Risky Mutable Proxy, and Unlimited Minting. AST features are derived via CodeBERT; CFG features via T... | {
"cdate": 1757838049602,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025natlm,\ntitle={{NATLM}: Detecting Defects in {NFT} Smart Contracts Leveraging {LLM}},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0IceiDrfxI},\nnote={under review}\n}"
},
"abstract": {
"value": "Security issues are becoming increasingly significant with the rapid evolution of Non-fungible Tokens (NFTs). The potential defects in NFT smart contracts could lead to substantial financial losses if exploited. To tackle this issue, this paper presents a framework called NATLM (NFT Assistant LLM), to detect potential defects in NFT smart contracts. NATLM effectively identifies 4 common types of vulnerabilities in NFT smart contracts, including ERC-721 Reentrancy, Public Burn, Risky Mutable Proxy, and Unlimited Minting. Relying exclusively on large language models (LLMs) for defect detection can lead to a high false-positive rate. To improve it, NATLM integrates static analysis with LLMs, specifically Gemini Pro 1.5. Initially, NATLM employs static analysis to extract structural, syntactic, and execution flow information from the code, represented through Abstract Syntax Trees (AST) and Control Flow Graphs (CFG). These extracted features are then combined with vectors of known defect examples to create a matrix for input into the knowledge base. Subsequently, the feature vectors and code vectors of the analyzed contract are compared with the contents in the knowledge base. Finally, the deep semantic analysis capabilities of LLM are used to identify defects in NFTs. Experimental results indicate that NATLM analyzed 8,672 collected NFT smart contracts, achieving an F1 score of 88.94\\%, outperforming other baselines."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"NFT",
"LLM",
"Smart Contract",
"Semantic Analysis"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/bdddf86568d8764dd37257e29334ad87194789b7.pdf"
},
"primary_area": {
"value": "other topics in machine learning (i.e., none of the above)"
},
"submission_guidelines": null,
"supplementary_material": {
"value": "/attachment/33d9a3380a902ed5ddb639a6f6db176ac71a716a.zip"
},
"title": {
"value": "NATLM: Detecting Defects in NFT Smart Contracts Leveraging LLM"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0IceiDrfxI",
"id": "0IceiDrfxI",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission5038/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897999057,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission5038/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission5038/Authors"
]
} | |
2,026 | 0Iw52EDu82 | [
2,
4,
6,
6
] | [
{
"content": "This paper investigates the scaling law of fully sparsely-activated language models. They first conduct experiments to compare different activation functions, sparsification functions, and gradient estimation methods. Then, they use scaling law (the relationship between cross-entropy loss and trai... | {
"cdate": 1758204654491,
"content": {
"TLDR": {
"value": "In this work, we investigate the architecture and scaling laws for fully sparsely-activated models, where every activation in linear transformations is sparse."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025scaling,\ntitle={Scaling Laws for Fully Sparsely-Activated Large Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0Iw52EDu82},\nnote={under review}\n}"
},
"abstract": {
"value": "Scaling laws play a crucial role in understanding and optimizing Large Language Models (LLMs). While previous work on scaling laws has primarily focused on either fully dense models or models with sparse Mixture of Experts (MoE), our work investigates fully sparsely-activated models, where every activation in linear transformations is sparse. We derive scaling laws for these models through extensive experiments with varying model sizes, training token counts, and activation sparsity ratios. Our findings demonstrate that fully sparsely-activated LLMs exhibit favorable scaling properties: as the total model size increases, LLMs can maintain higher activation sparsity while the performance gap between sparsely-activated and dense models narrows. Notably, our scaling laws indicate that a sparsely-activated full-precision model with a 45.58% sparsity ratio achieves optimal performance while maintaining the same number of active parameters. Furthermore, our scaling laws remain applicable to 1-bit pre-training of LLMs, suggesting promising directions for improving the efficiency of future models."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Activation sparsity",
"scaling law",
"large language models"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/68d8801b255d9afe13fd2fa81741bcc86a829bc4.pdf"
},
"primary_area": {
"value": "foundation or frontier models, including LLMs"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Scaling Laws for Fully Sparsely-Activated Large Language Models"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0Iw52EDu82",
"id": "0Iw52EDu82",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897545966,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission11921/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission11921/Authors"
]
} | |
2,026 | 0IwSQsqMU9 | [
8,
4,
4
] | [
{
"content": "Quite interesting work; A novel Darwinian perspective to optimization dynamics in NN. The paper presents a novel bio-inspired optimization method called Natural Selection (NS) that introduces explicit competition among training samples. By computing competitive scores through image stitching and d... | {
"cdate": 1757926747886,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025darwinian,\ntitle={Darwinian Optimization: Training Deep Networks with Natural Selection},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0IwSQsqMU9},\nnote={under review}\n}"
},
"abstract": {
"value": "In conventional deep learning training paradigms, all samples are usually subjected to uniform selective pressure, which fails to adequately account for variations in competitive intensity and diversity among them. This often leads to challenges such as class imbalance bias, insufficient learning of hard samples, and improper handling of noisy samples. Drawing inspiration from the principles of species competition and adaptation in natural ecosystems, we propose a bio-inspired optimization method for deep networks, termed Natural Selection (NS). NS introduces a competition mechanism by stitching and scaling a group of samples before forward prediction. Each sample is then assigned a natural selection score based on its prediction, reflecting its competitive status within the group. This score is further used to dynamically adjust the loss weight of each sample, thereby forming an optimization process that more closely mimics a Darwinian ecological equilibrium. Experimental results on 12 public datasets consistently demonstrate that NS improves performance without being tied to specific network architectures or task assumptions. This study offers a novel perspective on deep network optimization and holds instructive significance for broader applications. The code will be made publicly accessible."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Deep network optimization",
"natural selection",
"sample weighting",
"image classification",
"emotion recognition"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/bcff301a2f37f95db217c28b899c4e2e5e423b9d.pdf"
},
"primary_area": {
"value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Darwinian Optimization: Training Deep Networks with Natural Selection"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0IwSQsqMU9",
"id": "0IwSQsqMU9",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission5674/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897961688,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission5674/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission5674/Authors"
]
} | |
2,026 | 0JLUFJMo5p | [
0,
2,
0,
2
] | [
{
"content": "The manuscript strongly resembles AI-generated content and may have been produced as an internal test for prospective AI researchers. If so, it suggests that the current state of such roles remains immature and requires further development.",
"id": "nKMdFIiiCf",
"rating": 0
},
{
"c... | {
"cdate": 1758368192077,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025dynamic,\ntitle={Dynamic Task-Embedded Reward Machines for {\\textbackslash}{\\textbackslash} Adaptive Code Generation and Manipulation {\\textbackslash}{\\textbackslash} in Reinforcement Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0JLUFJMo5p},\nnote={under review}\n}"
},
"abstract": {
"value": "We introduce Dynamic Task-Embedded Reward Machine (DTERM), a new machine learning approach for reinforcement learning on tasks of code generation and code manipulation. Conventional reward models tend to be based on fixed weightings or manual tuning, which is not flexible enough for many different coding tasks, such as translation, completion and repair. To overcome that, DTERM dynamically modulates reward components using a hypernetwork-driven architecture, which can balance the task-aware configuration of syntactic correctness, semantic correctness, and computational efficiency. The framework combines three key modules, including a transformer-based task embedding generator, a modular reward decomposer, and a hypernetwork to generate context-dependent weights of sub-rewards."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Reinforcement Learning"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/fa6de8f172967f9988c29abcc16091879272bcd0.pdf"
},
"primary_area": {
"value": "transfer learning, meta learning, and lifelong learning"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Dynamic Task-Embedded Reward Machines for \\\\ Adaptive Code Generation and Manipulation \\\\ in Reinforcement Learning"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0JLUFJMo5p",
"id": "0JLUFJMo5p",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission25449/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759896720791,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission25449/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission25449/Authors"
]
} | |
2,026 | 0JWhSwwXak | [
4,
4,
6,
6
] | [
{
"content": "This paper proposes SYMMATIKA, a symbolic regression framework that combines multi-island genetic programming and a reusable symbol library to accelerate search, supporting both explicit (y=f(x)) and implicit (F(x,y)=0) regression tasks. Experimental results demonstrate its superiority over existi... | {
"cdate": 1758226660875,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025symmatika,\ntitle={SymMatika: Structure-Aware Symbolic Discovery},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0JWhSwwXak},\nnote={under review}\n}"
},
"abstract": {
"value": "Symbolic regression (SR) seeks to recover closed-form mathematical expressions that describe observed data. While existing methods have advanced the discovery of either explicit mappings (i.e., $y = f(\\mathbf{x})$) or discovering implicit relations (i.e., $F(\\mathbf{x}, y)=0$), few modern and accessible frameworks support both. Moreover, most approaches treat each expression candidate in isolation, without reusing recurring structural patterns that could accelerate search. We introduce SymMatika, a hybrid SR algorithm that combines multi-island genetic programming (GP) with a reusable motif library inspired by biological sequence analysis. SymMatika identifies high-impact substructures in top-performing candidates and reintroduces them to guide future generations. Additionally, it incorporates a feedback-driven evolutionary engine and supports both explicit and implicit relation discovery using implicit-derivative metrics. Across benchmarks, SymMatika achieves state-of-the-art recovery rates on the Nguyen and Feynman benchmark suites, an impressive recovery rate of 61\\% on Nguyen-12 compared to the next best 2\\%, and strong placement on the error-complexity Pareto fronts on the Feynman equations and on a subset of 57 SRBench Black-box problems. Our results demonstrate the power of structure-aware evolutionary search for scientific discovery. To support broader research in interpretable modeling and symbolic discovery, we have open-sourced the full SymMatika framework."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"AI for science",
"symbolic regression",
"genetic programming"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/5d2a444c0a7b4c93be4abcd3f68c1c25286e4c1c.pdf"
},
"primary_area": {
"value": "applications to physical sciences (physics, chemistry, biology, etc.)"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "SymMatika: Structure-Aware Symbolic Discovery"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0JWhSwwXak",
"id": "0JWhSwwXak",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission13998/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897397261,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission13998/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission13998/Authors"
]
} | |
2,026 | 0JYtXNl7ns | [
2,
2,
4,
4
] | [
{
"content": "The paper introduces an inference-time scaling framework (SHARS), aiming to allocate additional computational resources to detect and mitigate hallucinations during decoding.\nAs a main component of the framework, the uncertainty-based hallucination detection method HalluSE which aims to improve s... | {
"cdate": 1758288192041,
"content": {
"TLDR": {
"value": "an inference-time scaling framework for hallucination mitigation in open-ended generation."
},
"_bibtex": {
"value": "@inproceedings{\nanonymous2025building,\ntitle={Building Reliable Long-Form Generation via Step-Wise Hallucination Rejection Sampling},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0JYtXNl7ns},\nnote={under review}\n}"
},
"abstract": {
"value": "Large language models (LLMs) have achieved remarkable progress in open-ended text generation, yet they remain prone to hallucinating incorrect or unsupported content, which undermines their reliability. This issue is exacerbated in long-form generation due to hallucination snowballing, a phenomenon where early errors propagate and compound into subsequent outputs. To address this challenge, we propose a novel inference-time scaling framework, named Step-wise HAllucination Rejection Sampling (SHARS), that allocates additional computation during decoding to detect and reject hallucinated content as it is produced. By retaining only confident information and building subsequent generations upon it, the framework mitigates hallucination accumulation and enhances factual consistency. To instantiate this framework, we further introduce a new uncertainty-based hallucination detection method, named HalluSE, for long-form generation, improving upon the prior semantic entropy approach. The combined system enables models to self-correct hallucinations without requiring external resources such as web search or knowledge bases, while remaining compatible with them for future extensions. Empirical evaluations on standardized hallucination benchmarks demonstrate that our method substantially reduces hallucinations in long-form generation while preserving or even improving the informativeness of generation."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"hallucination",
"inferece-time scaling",
"large language models",
"semantic entropy"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/4c0a73711949f5a4c0e597230eb920ffb943bbf3.pdf"
},
"primary_area": {
"value": "alignment, fairness, safety, privacy, and societal considerations"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "Building Reliable Long-Form Generation via Step-Wise Hallucination Rejection Sampling"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0JYtXNl7ns",
"id": "0JYtXNl7ns",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission18485/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897100328,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission18485/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission18485/Authors"
]
} | |
2,026 | 0JayjvOKxt | [
2,
4,
6,
4
] | [
{
"content": "This paper proposes the Adaptive and Selective Reset (ASR) scheme to address the problem of model collapse in long-term Test-Time Adaptation (TTA). The main contributions are: 1) The ASR mechanism dynamically determines when and which parts of the model to reset; 2) An importance-aware knowledge r... | {
"cdate": 1758270901913,
"content": {
"TLDR": null,
"_bibtex": {
"value": "@inproceedings{\nanonymous2025when,\ntitle={When and Where to Reset Matters for Long-Term Test-Time Adaptation},\nauthor={Anonymous},\nbooktitle={Submitted to The Fourteenth International Conference on Learning Representations},\nyear={2025},\nurl={https://openreview.net/forum?id=0JayjvOKxt},\nnote={under review}\n}"
},
"abstract": {
"value": "When continual test-time adaptation (TTA) persists over the long term, errors accumulate in a model and further lead it to predict only a few classes regardless of the input, known as model collapse. Recent studies have explored reset strategies that erase these accumulated errors completely. However, their periodic resets lead to suboptimal adaptation, as they occur independently of collapse. Also, their full resets cause the catastrophic loss of knowledge acquired over time, even though it could be beneficial in future. To this end, we propose 1) an Adaptive and Selective Reset (ASR) scheme that dynamically determines when and where to reset, 2) an importance-aware regularizer to recover essential knowledge lost from reset, and 3) an on-the-fly adaptation adjustment scheme to enhance adaptability under challenging domain shifts. Extensive experiments across long-term TTA benchmarks demonstrate the effectiveness of our approach, particularly under challenging conditions. Our code will be released."
},
"anonymous_url": null,
"authorids": null,
"authors": null,
"code_of_ethics": null,
"keywords": {
"value": [
"Test-Time Adaptation",
"Continual Test-Time Adaptation"
]
},
"no_acknowledgement_section": null,
"paperhash": null,
"pdf": {
"value": "/pdf/504c73f24c6ccca26313babd1bdf7dfd05964f6b.pdf"
},
"primary_area": {
"value": "transfer learning, meta learning, and lifelong learning"
},
"submission_guidelines": null,
"supplementary_material": null,
"title": {
"value": "When and Where to Reset Matters for Long-Term Test-Time Adaptation"
},
"venue": {
"value": "ICLR 2026 Conference Submission"
},
"venueid": {
"value": "ICLR.cc/2026/Conference/Submission"
}
},
"forum": "0JayjvOKxt",
"id": "0JayjvOKxt",
"invitations": [
"ICLR.cc/2026/Conference/-/Submission",
"ICLR.cc/2026/Conference/-/Post_Submission",
"ICLR.cc/2026/Conference/Submission16981/-/Full_Submission"
],
"license": "CC BY 4.0",
"mdate": 1759897206393,
"odate": 1759896705795,
"readers": [
"everyone"
],
"signatures": [
"ICLR.cc/2026/Conference/Submission16981/Authors"
],
"writers": [
"ICLR.cc/2026/Conference",
"ICLR.cc/2026/Conference/Submission16981/Authors"
]
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.