arXivDaily arXiv每日学术速递 周一至周五更新

期刊&会议

International Conference on Learning Representations · 会议 · Machine Learning

2026-06-09 至 2026-06-09 共收录 10
2606.09758 2026-06-09 cs.RO cs.AI cs.LG 新提交

Difference-Aware Retrieval Policies for Imitation Learning

差异感知的模仿学习检索策略

Quinn Pfeifer, Ethan Pronovost, Paarth Shah, Khimya Khetarpal, Siddhartha Srinivasa, Abhishek Gupta

机构 * Paul G. Allen School of Computer Science & Engineering, University of Washington(华盛顿大学保罗·G·艾伦计算机科学与工程学院) Toyota Research Institute(丰田研究所) Google DeepMind(谷歌DeepMind) Mila

AI总结 提出DARP,一种半参数检索式模仿学习方法,通过基于k近邻的局部邻域结构重参数化,解决行为克隆的分布外泛化问题,在连续控制和机器人操作任务中性能提升15-46%。

Comments 12 pages, 7 figures, 3 tables. Accepted to ICLR 2026. Code and demos available at https://weirdlabuw.github.io/darp-site/

详情

展开后加载摘要…

URL PDF HTML 收藏
2606.09401 2026-06-09 cs.LG cs.CR 新提交

Benchmarking Empirical Privacy Protection for Adaptations of Large Language Models

大语言模型适配的实证隐私保护基准测试

Bartłomiej Marek, Lorenzo Rossi, Vincent Hanke, Xun Wang, Michael Backes, Franziska Boenisch, Adam Dziedzic

机构 * CISPA Helmholtz Center for Information Security(CISPA 欧洲信息安全中心)

AI总结 通过系统变化适配数据分布,使用鲁棒成员推断和金丝雀数据提取攻击,评估差分隐私下大语言模型的实际隐私风险,发现分布偏移显著影响隐私脆弱性,LoRA等参数高效微调方法对分布外数据提供最佳实证保护。

Comments Accepted at ICLR 2026 (Oral)

详情

展开后加载摘要…

URL PDF HTML 收藏
2606.09257 2026-06-09 cs.LG cs.AI stat.ML 新提交

BSTabDiff: Block-Subunit Diffusion Priors for High-Dimensional Tabular Data Generation

BSTabDiff: 用于高维表格数据生成的块-子单元扩散先验

Al Zadid Sultan Bin Habib, Md Younus Ahamed, Prashnna Gyawali, Gianfranco Doretto, Donald A. Adjeroh

机构 * West Virginia University(西弗吉尼亚大学) The University of Utah(犹他大学)

AI总结 针对高维低样本量表格数据,提出BSTabDiff框架,通过将特征划分为潜在块并使用共享低维子单元变量生成每个块,结合扩散先验和copula依赖,实现稳定合成与可控基准生成。

Comments Published as a paper at the 2nd DeLTa Workshop, ICLR 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2606.08578 2026-06-09 cs.LG 新提交

Lost in the Non-convex Loss Landscape: How to Fine-tune the Large Time Series Model?

迷失在非凸损失景观中:如何微调大型时间序列模型?

Xu Zhang, Peang Wang, Wei Wang

机构 * Shanghai Key Laboratory of Data Science(上海市数据科学重点实验室) College of Computer Science and Artificial Intelligence(计算机科学与人工智能学院) Fudan University(复旦大学)

AI总结 针对预训练大型时间序列模型微调时因非凸损失景观导致过拟合的问题,提出平滑全微调(SFF)方法,通过随机初始化辅助模型插值平滑损失景观,提升可训练性,在八个代表性模型上取得一致改进。

Comments This paper has been accepted by The Fourteenth International Conference on Learning Representations (ICLR 2026). The code is available at the link \url{https://github.com/Meteor-Stars/SFF}

详情

展开后加载摘要…

URL PDF HTML 收藏
2606.08574 2026-06-09 cs.LG cs.CV 新提交

OrderDP: A Theoretically Guaranteed Lossless Dynamic Data Pruning Framework

OrderDP:一种理论上保证无损的动态数据剪枝框架

Chenhan Jin, Shengze Xu, Qingsong Wang, Fan Jia, Dingshuo Chen, Tieyong Zeng

机构 * The Chinese University of Hong Kong(香港中文大学) Beijing Normal-Hong Kong Baptist University(北京师范大学-香港 Baptist大学) Guangzhou Nanfang College(广州南方学院) Institute of Automation, Chinese Academy of Sciences(中国科学院自动化研究所) Xiangtan University(湘潭大学) University of Utah(犹他大学)

AI总结 提出OrderDP框架,通过随机子集选取与top-q样本选择实现无偏梯度估计,提供收敛性和泛化性理论保证,在CIFAR和ImageNet上降低40%训练成本且保持精度。

Comments Published as a conference paper at ICLR 2026

Journal ref International Conference on Learning Representations (ICLR), 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2606.07728 2026-06-09 cs.LG 新提交

Characterizing the Discrete Geometry of ReLU Networks

表征ReLU网络的离散几何

Blake B. Gaines, Jinbo Bi

机构 * University of Connecticut(康涅狄格大学)

AI总结 本文研究全连接ReLU网络线性区域构成的复形,证明其连通图平均度上界为输入维度的两倍,且直径上界与输入维度无关。

Comments Selected for an oral presentation at ICLR 2026. Tagged PDF, reviews, and discussions are available at https://openreview.net/forum?id=TgLW2DiRDG

Journal ref Proceedings of the International Conference on Learning Representations (ICLR), 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2604.09787 2026-06-09 astro-ph.IM astro-ph.GA cs.LG 版本更新

Learning What's Real: Disentangling Signal and Measurement Artifacts in Multi-Sensor Data, with Applications to Astrophysics

学习真实内容:在多传感器数据中分离信号和测量伪影,应用于天体物理学

Pablo Mercader-Perez, Carolina Cuesta-Lazaro, Daniel Muthukrishna, Jeroen Audenaert, V. Ashley Villar, David W. Hogg, Marc Huertas-Company, William T. Freeman

机构 * Massachusetts Institute of Technology(麻省理工学院) Flatiron Institute, Simons Foundation(Flatiron研究所,Simons基金会) Institute for Advanced Studies(高级研究 institute) Harvard University(哈佛大学) New York University(纽约大学) Instituto de Astrofísica de Canarias(加那利大天文台)

AI总结 本文提出一种深度学习框架,通过重叠观测、双编码器架构和反事实生成目标,分离多传感器数据中的信号与伪影,提升天体物理学研究的准确性。

Comments Accepted at the 2nd Workshop on Foundation Models for Science at ICLR 2026. 10 pages, 7 figures (main text), plus appendix

详情

展开后加载摘要…

URL PDF HTML 收藏
2604.07848 2026-06-09 cs.LG q-bio.MN 版本更新

Information-Theoretic Requirements for Gradient-Based Task Affinity Estimation in Multi-Task Learning

基于梯度的任务亲和性估计在多任务学习中的信息论要求

Jasper Zhang, Bryan Cheng

机构 * Great Neck South High School(Great Neck South 高中)

AI总结 本文探讨了多任务学习中梯度基于任务亲和性估计的信息论要求,发现任务样本重叠度对梯度对齐的影响,并揭示了样本重叠度的相变特性。

Comments 8 pages, 4 figures. ACM BCB 2026 Short Paper. Accepted at workshop on AI for Accelerated Materials Design, Foundation Models for Science: Real-World Impact and Science-First Design, and Generative and Experimental Perspectives for Biomolecular Design at ICLR 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2509.24762 2026-06-09 cs.LG 版本更新

In-Context Learning of Temporal Point Processes with Foundation Inference Models

基于基础推理模型的时间点过程上下文学习

David Berghaus, Patrick Seifner, Kostadin Cvejoski, César Ojeda, Ramsés J. Sánchez

机构 * Lamarr Institute(拉马尔研究所) Fraunhofer IAIS(弗劳恩霍夫人工智能研究所) University of Bonn(波恩大学) JetBrains Research(JetBrains研究) University of Potsdam(波恩大学)

AI总结 提出一种基于摊销推理和上下文学习的点过程基础推理模型FIM-PP,通过大规模合成数据预训练,无需额外训练即可估计真实MTPP,或快速微调至目标系统。

Comments This paper is published as a conference paper at ICLR 2026

Journal ref The Fourteenth International Conference on Learning Representations (ICLR 2026)

详情

展开后加载摘要…

URL PDF HTML 收藏
2510.17947 2026-06-09 cs.CR cs.AI cs.CL cs.LG cs.MA 版本更新

PLAGUE: Plug-and-play framework for Lifelong Adaptive Generation of Multi-turn Exploits

PLAGUE:面向多轮利用的终身自适应生成的即插即用框架

Neeladri Bhuiya, Madhav Aggarwal, Diptanshu Purwar

机构 * A10 Networks, Inc.(A10网络公司) University of Massachusetts Amherst(马萨诸塞大学阿姆赫斯特分校)

AI总结 提出PLAGUE框架,通过终身学习启发的三阶段设计(Primer、Planner、Finisher)实现高效多轮越狱攻击,在o3和Opus 4.1等强安全模型上ASR提升超30%。

Comments Accepted in ICLR 2026

详情

展开后加载摘要…

URL PDF HTML 收藏