Transformers Provably Learn Chain-of-Thought Reasoning with Length Generalization
Comments This is the full version of a paper published at NeurIPS 2025
期刊&会议
NeurIPS
Conference on Neural Information Processing Systems · 会议 · Machine Learning
Self-Evaluating LLMs for Multi-Step Tasks: Stepwise Confidence Estimation for Failure Detection
机构 * Dyania Health（Dyania健康）
Comments Accepted at NeurIPS 2025 Workshop on Evaluating the Evolving LLM Lifecycle: Benchmarks, Emergent Abilities, and Scaling
Inference-Time Scaling of Diffusion Models for Infrared Data Generation
机构 * Cornell University（康奈尔大学） ； YRIKKA, Inc.（YRIKKA公司）
Comments Peer-reviewed workshop paper
Journal ref 39th Conference on Neural Information Processing Systems (NeurIPS 2025) Workshop: Learning to Sense
E2E-VGuard: Adversarial Prevention for Production LLM-based End-To-End Speech Synthesis
机构 * Shenzhen International Graduate School, Tsinghua University（清华大学深圳国际研究生院） ； Beijing University of Posts and Telecommunications（北京邮电大学） ； CSIRO’s Data61（CSIRO数据61） ； Responsible AI Research (RAIR) Centre, The University of Adelaide（阿德莱德大学负责任人工智能研究（RAIR）中心） ； National University of Singapore（新加坡国立大学） ； The University of Hong Kong（香港大学）
Comments Accepted to NeurIPS 2025
Breaking the Gradient Barrier: Unveiling Large Language Models for Strategic Classification
机构 * College of Computer Science and Technology, National University of Defense Technology（国防科技大学计算机科学与技术学院） ； Center for Data Science, Peking University（北京大学数据科学中心） ； Faculty of Engineering, the University of Sydney（悉尼大学工程学院） ； Faculty of Computing, Harbin Institute of Technology（哈尔滨工业大学计算机学院）
Comments Accepted by NeurIPS 2025
Hybrid Autoencoders for Tabular Data: Leveraging Model-Based Augmentation in Low-Label Settings
机构 * Bar-Ilan University（巴伊兰大学）
Comments accepted to neurips 2025, main text is 10 pages
Learning to Focus: Prioritizing Informative Histories with Structured Attention Mechanisms in Partially Observable Reinforcement Learning
机构 * Delft University of Technology（代尔夫特理工大学）
Comments Accepted to Embodied World Models for Decision Making (EWM) Workshop at NeurIPS 2025
Gaussian-Augmented Physics Simulation and System Identification with Complex Colliders
机构 * Istituto Italiano di Tecnologia（意大利技术研究院） ； UC San Diego（加州大学圣地亚哥分校）
Comments Accepted to NeurIPS 2025. Project website: https://as-diffmpm.github.io/
Coupling Agent-based Modeling and Life Cycle Assessment to Analyze Trade-offs in Resilient Energy Transitions
机构 * Lawrence Berkeley National Laboratory（伯克利劳伦斯国家实验室）
Comments 4 pages (+4 pages in appendix), 3 figures (+ 2 figures in appendix), 8 tables in appendix, NeurIPS Workshop on Tackling Climate Change with Machine Learning, 2025
Rank-1 LoRAs Encode Interpretable Reasoning Signals
机构 * MATS
Comments 39th Conference on Neural Information Processing Systems (NeurIPS 2025) Workshop: Mechanistic Interpretability Workshop
Optimistic Online-to-Batch Conversions for Accelerated Convergence and Universality
机构 * National Key Laboratory for Novel Software Technology, Nanjing University, China（新型软件技术国家实验室，南京大学，中国） ； School of Artificial Intelligence, Nanjing University, China（人工智能学院，南京大学，中国）
Comments NeurIPS 2025
FractalBench: Diagnosing Visual-Mathematical Reasoning Through Recursive Program Synthesis
机构 * MIT（麻省理工学院） ； Comenius University in Bratislava（布拉迪斯拉发孔院大学）
Comments Accepted to The 5th Workshop on Mathematical Reasoning and AI at the 39th Conference on Neural Information Processing Systems (NeurIPS 2025); 25 pages, 14 figures, 8 tables; Code available at https://github.com/NaiveNeuron/FractalBench
ConnectomeBench: Can LLMs Proofread the Connectome?
机构 * MIT（麻省理工学院） ； HHMI（霍华德·休斯医学研究所） ； McGovern Institute（麦戈文研究所） ； MIT Departments of Brain and Cognitive Sciences（麻省理工学院脑科学与认知科学系）
Comments To appear in NeurIPS 2025 Datasets and Benchmarks Track
Matchings Under Biased and Correlated Evaluations
Comments To appear in NeurIPS 2025
Bridging the gap to real-world language-grounded visual concept learning
机构 * School of Computing, KAIST（计算机学院，韩国科学技术院）
Journal ref Advances in Neural Information Processing Systems (NeurIPS), 2025
Universal Spectral Tokenization via Self-Supervised Panchromatic Representation Learning
Comments Accepted at NeurIPS 2025 Machine Learning and the Physical Sciences Workshop; v2: added collaboration
Lagrangian neural ODEs: Measuring the existence of a Lagrangian with Helmholtz metrics
机构 * Interdisciplinary Center for Scientific Computing（交叉科学计算中心） ； Heidelberg University（海德堡大学） ； Astronomisches Rechen-Institut der Universität Heidelberg（海德堡大学天文计算研究所）
Comments Accepted for the NeurIPS 2025 Machine Learning and the Physical Sciences workshop. 6 pages, 3 figures
UniPixel: Unified Object Referring and Segmentation for Pixel-Level Visual Reasoning
机构 * The Hong Kong Polytechnic University（香港理工大学） ； ARC Lab, Tencent PCG（腾讯PCG ARC实验室） ； Institute of Automation, Chinese Academy of Sciences（中国科学院自动化研究所） ； vivo Mobile Communication Co.（vivo移动通信公司） ； MindWingman Technology (Shenzhen) Co., Ltd.（深圳MindWingman技术有限公司）
Comments NeurIPS 2025 Camera Ready. Project Page: https://polyu-chenlab.github.io/unipixel/
Visual Structures Helps Visual Reasoning: Addressing the Binding Problem in VLMs
Comments Accepted to NeurIPS 2025 (Thirty-ninth Conference on Neural Information Processing Systems)
Consistent Story Generation: Unlocking the Potential of Zigzag Sampling
机构 * KU Leuven（库勒芬大学） ； Utrecht University（乌得勒支大学）
Comments 20 pages, 10 figures
Journal ref published at NeurIPS 2025
Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion
机构 * Adobe Research（Adobe研究院） ； The University of Texas at Austin（德克萨斯大学奥斯汀分校）
Comments NeurIPS 2025 spotlight. Project website: http://self-forcing.github.io/
DeepDiver: Adaptive Search Intensity Scaling via Open-Web Reinforcement Learning
机构 * Huawei Language Model Lab（华为语言模型实验室）
Comments Accepted as NeurIPS 2025 Spotlight
Robust Hallucination Detection in LLMs via Adaptive Token Selection
机构 * Imperial College London, UK（伦敦帝国学院） ； Singapore Management University（新加坡管理大学）
Comments Accepted by NeurIPS 2025
Learning Temporal 3D Semantic Scene Completion via Optical Flow Guidance
机构 * College of Computer Science and Electronic Engineering, Hunan University（湖南大学计算机科学与电子工程学院） ； State Key Laboratory of Advanced Design and Manufacturing Technology for Vehicle, Hunan University（车辆先进设计与制造技术国家重点实验室）
Comments Accepted by NeurIPS 2025
Smooth Sailing: Lipschitz-Driven Uncertainty Quantification for Spatial Association
机构 * MIT LIDS（麻省理工学院LIDS）
Comments The first two authors contributed equally; 39 pages, 14 figures
Journal ref NeurIPS 2025
Mitigating Sexual Content Generation via Embedding Distortion in Text-conditioned Diffusion Models
机构 * Department of Artificial Intelligence（人工智能系） ； Kyungpook National University（全北国立大学）
Comments NeurIPS 2025 accepted. Official code: https://github.com/amoeba04/des
GC4NC: A Benchmark Framework for Graph Condensation on Node Classification with New Insights
机构 * Emory University（埃默里大学） ； Google Deepmind（谷歌DeepMind）
Comments 30 pages, Accepted by NeurIPS 2025 Datasets & Benchmarks Track
Reconstruction and Secrecy under Approximate Distance Queries
机构 * Departments of Mathematics, Computer Science, and Data and Decision Sciences, Technion and Google Research（技术学院数学、计算机科学和数据与决策科学系，以及谷歌研究） ； Departments of Mathematics, Technion（技术学院数学系）
Comments 39 pages. Conference version: NeurIPS 2025 (Spotlight). Extended appendix included
Vocabulary In-Context Learning in Transformers: Benefits of Positional Encoding
机构 * School of Mathematical Sciences, Beijing Normal University（北京师范大学数学学院）
Comments Accepted as NIPS 2025 poster
Decomate: Leveraging Generative Models for Co-Creative SVG Animation
机构 * MODULABS
Comments Accepted at the 1st Workshop on Generative and Protective AI for Content Creation (NeurIPS 2025)
