arXivDaily arXiv每日学术速递 周一至周五更新

期刊&会议

International Conference on Machine Learning · 会议 · Machine Learning

2026-06-12 至 2026-06-12 共收录 16
2605.00600 2026-06-12 cs.LG cs.AI cs.CV 版本更新

Possibilistic Predictive Uncertainty for Deep Learning

深度学习的可能性预测不确定性

Yao Ni, Jeremie Houssineau, Yew-Soon Ong, Piotr Koniusz

机构 * University of Cambridge(剑桥大学) National University of Singapore(新加坡国立大学) University of Warsaw(华沙大学)

AI总结 提出基于可能性理论的Dirichlet近似可能性后验预测(DAPPr)框架,通过投影-近似策略实现高效且原则性的认知不确定性量化,在多个基准上达到竞争性能。

Comments Accepted by ICML 2026, 20 pages

详情

展开后加载摘要…

URL PDF HTML 收藏
2604.26940 2026-06-12 cs.CL 版本更新

Select to Think: Unlocking SLM Potential with Local Sufficiency

Select to Think: 利用局部充分性解锁小语言模型潜力

Wenxuan Ye, Yangyang Zhang, Xueli An, Georg Carle, Yunpu Ma

机构 * University of Science and Technology of China(中国科学技术大学)

AI总结 提出Select to Think (S2T)方法,通过将大语言模型角色从生成转为选择,并蒸馏选择逻辑到小语言模型,使其在推理时无需依赖大模型,显著提升性能。

Comments Accepted to ICML 2026. Code is available at https://github.com/YeRona/Select-to-Think

详情

展开后加载摘要…

URL PDF HTML 收藏
2603.14407 2026-06-12 cs.LG 版本更新

Towards One-for-All Anomaly Detection for Tabular Data

面向表格数据的通用异常检测

Shiyuan Li, Yixin Liu, Yu Zheng, Xiaofeng Cao, Shirui Pan, Heng Tao Shen

机构 * University of Science and Technology of China(中国科学技术大学)

AI总结 提出OFA-TAD框架,通过多视图邻居距离表示和混合专家评分网络,实现跨领域表格异常检测的通用化,一次训练即可泛化到未见数据集。

Comments Accepted by ICML 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2603.00610 2026-06-12 cs.SD cs.AI cs.LG cs.MM eess.AS 版本更新

CMI-RewardBench: Evaluating Music Reward Models with Compositional Multimodal Instruction

CMI-RewardBench: 基于组合多模态指令评估音乐奖励模型

Yinghao Ma, Haiwen Xia, Hewei Gao, Weixiong Chen, Yuxin Ye, Yuchen Yang, Sungkyun Chang, Mingshuo Ding, Yizhi Li, Ruibin Yuan, Simon Dixon, Emmanouil Benetos

机构 * National University of Singapore(新加坡国立大学) University of Science and Technology of China(中国科学技术大学) University of Cambridge(剑桥大学) University of Toronto(多伦多大学)

AI总结 针对音乐生成模型缺乏有效评估机制的问题,提出CMI-RewardBench基准,包含大规模偏好数据集和参数高效奖励模型,实现多模态指令下的音乐质量评估。

Comments Accepted by ICML 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2510.02524 2026-06-12 cs.CL cs.FL cs.LG 版本更新

Unraveling Syntax: Language Modeling and the Substructure of Grammars

解析句法:语言建模与语法的子结构

Laura Ying Schulz, Daniel Mitropolsky, Tomaso Poggio

机构 * Massachusetts Institute of Technology(麻省理工学院)

AI总结 本文研究语言模型在上下文无关语法子结构上的学习行为,证明损失函数在顶层子语法上线性递归,并发现参数化模型并行学习子语法,子语法预训练能提升小模型性能并改善内部表征。

Comments Equal contribution by LYS and DM. Accepted to the 43rd International Conference on Machine Learning (ICML 2026)

详情

展开后加载摘要…

URL PDF HTML 收藏
2602.14367 2026-06-12 cs.CL cs.AI cs.IR cs.LG 版本更新

InnoEval: On Research Idea Evaluation as a Knowledge-Grounded, Multi-Perspective Reasoning Problem

InnoEval:将研究思路评估视为基于知识的多视角推理问题

Shuofei Qiao, Yunxiang Wei, Xuehai Wang, Bin Wu, Boyang Xue, Ningyu Zhang, Hossein A. Rahmani, Yanshan Wang, Qiang Zhang, Keyan Ding, Jeff Z. Pan, Huajun Chen, Emine Yilmaz

机构 * University of Science and Technology of China(中国科学技术大学)

AI总结 提出InnoEval框架,通过异构深度知识检索和多视角评审委员会,实现基于知识的多维度解耦评估,在点对点、成对和分组评估任务中优于基线方法。

Comments ICML 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2602.04208 2026-06-12 cs.RO cs.AI cs.LG 版本更新

SCALE: Self-uncertainty Conditioned Adaptive Looking and Execution for Vision-Language-Action Models

SCALE: 基于自不确定性条件自适应观察与执行的视觉-语言-动作模型

Hyeonbeom Choi, Daechul Ahn, Youhan Lee, Taewook Kang, Seongwon Cho, Jonghyun Choi

机构 * Seoul National University(首尔国立大学)

AI总结 提出SCALE推理策略,利用自不确定性联合调节视觉感知和动作,无需额外训练或验证器,仅单次前向传播,提升VLA模型在模拟和真实环境中的鲁棒性。

Comments ICML 2026 Spotlight. Project page: https://dcahn12.github.io/projects/scale/

详情

展开后加载摘要…

URL PDF HTML 收藏
2601.09693 2026-06-12 cs.LG stat.ML 版本更新

Contrastive Geometric Learning Unlocks Unified Structure- and Ligand-Based Drug Design

对比几何学习实现统一的结构与配体药物设计

Lisa Schneckenreiter, Sohvi Luukkonen, Lukas Friedrich, Daniel Kuhn, Günter Klambauer

机构 * DeepMind Ltd(DeepMind有限公司)

AI总结 提出对比几何模型ConGLUDe,统一结构与配体训练,实现虚拟筛选、靶标钓鱼和配体条件口袋预测,在多项基准测试中表现优异。

Comments Forty-Third International Conference on Machine Learning

详情

展开后加载摘要…

URL PDF HTML 收藏
2512.12571 2026-06-12 cs.CV 版本更新

Measurement Plasticity: Sensor-Level Adaptation for Vision-Language Models

测量塑性:面向视觉-语言模型的传感器级自适应

Boyeong Im, Wooseok Lee, Yoojin Kwon, Hyung-Sin Kim

机构 * University of Seoul(首尔大学)

AI总结 提出多视角物理提示(MVP)用于测试时自适应,通过将相机曝光三角(ISO、快门速度、光圈)作为物理提示,在传感器层面进行自适应,无需梯度或模型修改,在ImageNet-ES上优于数字方法。

Comments Accepted to the ICML 2026 Workshop on Continual Adaptation at Scale

详情

展开后加载摘要…

URL PDF HTML 收藏
2601.22594 2026-06-12 cs.CL cs.AI 版本更新

Language Model Circuits Are Sparse in the Neuron Basis

语言模型电路在神经元基上是稀疏的

Aryaman Arora, Zhengxuan Wu, Jacob Steinhardt, Sarah Schwettmann

机构 * Stanford University(斯坦福大学)

AI总结 本文实证发现MLP神经元与稀疏自编码器一样是稀疏特征基,并基于此开发了端到端梯度归因流水线,在多项任务中揭示了因果有效的神经元电路。

Comments ICML Spotlight, camera-ready

详情

展开后加载摘要…

URL PDF HTML 收藏
2601.22003 2026-06-12 stat.ML cs.LG stat.CO 版本更新

Efficient Stochastic Optimisation via Sequential Monte Carlo

通过序贯蒙特卡洛实现高效随机优化

James Cuin, Davide Carbone, Yanbo Tang, O. Deniz Akyildiz

机构 * University of California, Berkeley(加州大学伯克利分校)

AI总结 针对梯度难以计算的优化问题,提出用序贯蒙特卡洛(SMC)采样器替代昂贵的内采样循环,实现高效随机优化,并在能量模型奖励调优中验证有效性。

Comments Accepted to ICML 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2601.21324 2026-06-12 stat.ML cs.LG 版本更新

Bulk-Calibrated Credal Ambiguity Sets: Fast, Tractable Decision Making under Out-of-Sample Contamination

批量校准的置信模糊集:样本外污染下的快速、可处理决策

Mengqi Chen, Thomas B. Berrett, Theodoros Damoulas, Michele Caprio

机构 * University of Bristol(布里斯托大学) University of Cambridge(剑桥大学) University of California, Berkeley(加州大学伯克利分校) University of Oxford(牛津大学)

AI总结 提出批量校准置信模糊集,通过分离批量内污染和尾部贡献,得到闭式有限风险目标,转化为线性或二阶锥规划,实现高效鲁棒优化。

Comments Accepted for publication (spotlight) at ICML 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2509.18085 2026-06-12 cs.LG cs.AI cs.CL 版本更新

Structuring The Future: Diffusion LLM Speculative Decoding via Calibrated Draft Graphs

构建未来:通过校准草稿图实现扩散LLM推测解码

Sudhanshu Agrawal, Risheek Garrepalli, Raghavv Goel, Christopher Lott, Fatih Porikli, Mingu Lee

机构 * University of Waterloo(多伦多大学)

AI总结 提出Spiffy算法,利用校准的草稿图结构实现扩散LLM的推测解码,在保持输出分布的同时加速推理,最高减少8.6倍模型推理次数并加速6.3倍令牌生成速率。

Comments Original version uploaded on Sep 22, 2025. (v2): Extended Table 2 with additional analysis and referenced it in Sec 5.2. (v3): Added note to Sec 4.2 and Appendix A.2 specifying conditions for losslessness. (v4): Updated with the version accepted to ICML 2026 workshops

详情

展开后加载摘要…

URL PDF HTML 收藏
2304.13836 2026-06-12 cs.LG cs.AI cs.CV stat.ME 版本更新

On Pitfalls of $\textit{RemOve-And-Retrain}$: Data Processing Inequality Perspective

论 $\textit{RemOve-And-Retrain}$ 的陷阱:数据处理不等式视角

Junhwa Song, Keumgang Cha, Junghoon Seo

机构 * KAIST(韩国科学技术院)

AI总结 从信息论角度揭示ROAR基准的缺陷:数据无关的后处理可提升ROAR分数,导致对归因图信息量的误判,并发现模糊性偏差。

Comments Accepted at the 2026 ICML Workshop on Mechanistic Interpretability

详情

展开后加载摘要…

URL PDF HTML 收藏
2512.23566 2026-06-12 math.DS cond-mat.stat-mech cs.LG math.OC stat.ML 版本更新

From geometry to dynamics: Learning overdamped Langevin dynamics from sparse observations with geometric constraints

从几何到动力学:基于几何约束从稀疏观测学习过阻尼朗之万动力学

Dimitra Maoutsa

机构 * Dimitra Maoutsa(迪米特拉·马乌茨)

AI总结 提出一种随机控制框架,利用系统不变密度的几何结构进行路径增强,从稀疏时间采样数据中恢复过阻尼朗之万动力学,无需参数模型假设。

Comments 10+54 pages, 14 figures; accepted at ICML 2026 An earlier account of this work has previously appeared in arXiv:2301.08102 and arXiv:2304.00423 ; main methodology remains the same, this version includes additional numerical experiments and theory

详情

展开后加载摘要…

URL PDF HTML 收藏
2505.20076 2026-06-12 cs.LG 版本更新

ExPLAIND: Unifying Model, Data, and Training Attribution to Study Model Behavior

ExPLAIND:统一模型、数据和训练归因以研究模型行为

Florian Eichin, Yupei Du, Philipp Mondorf, Maria Matveev, Barbara Plank, Michael A. Hedderich

机构 * University of Michigan(密歇根大学)

AI总结 提出ExPLAIND框架,统一归因于模型组件、数据和训练轨迹,支持跨粒度解释,通过梯度路径核和AdamW核机器推导参数级和步骤级影响分数,验证了Transformer的Grokking和EuroLLM预训练中的两阶段动态。

Comments published at ICML 2026, code at https://github.com/mainlp/explaind

详情

展开后加载摘要…

URL PDF HTML 收藏