ICML

Bend the Basics: Degradation-Aware Deformable Tokenization for All-in-One Image Restoration

基础适配：面向全场景图像恢复的退化感知可变形分词

Zihao He, Yunfeng Wu, Xinchao Wang, Songhua Liu

AI总结 本文提出FIT模型，通过全流程退化感知设计及任务分词丢弃策略，在五组图像恢复基准上实现最优性能，较现有方法提升0.5~1.1 dB，可适配各类退化图像恢复任务。

Comments 16 pages, 12 figures. Accepted to ICML 2026

Journal ref Proceedings of the 43rd International Conference on Machine Learning (ICML 2026), PMLR 306, 2026

Faster Query-Key Learning Sharpens Attention in Self-Attention Models

更快的查询-键学习强化自注意力模型中的注意力机制

Rahul Vashisht, Harish G. Ramaswamy

AI总结 该研究通过分析自注意力模型的查询-键与输出-值回路的参数化，发现更快的查询-键学习可强化注意力、提升可解释性，且预测性能相当。

Comments Accepted to the 43rd International Conference on Machine Learning (ICML 2026)

Suppress and Diversify: Refining Robust Pathways for Corruption Robustness

抑制与多样化：优化针对自然图像损坏的鲁棒路径

Jiangang Yang, Wenhui Shi, Xiaoran Xu, Wenyue Chong, Luqing Luo, Jing Xing, Jian Liu

AI总结 该研究针对自然图像损坏的模型鲁棒性问题，提出非侵入式方法S&D，通过选择并多样化鲁棒路径提升性能，在多场景多任务中展现出广泛有效性。

Comments Accepted at ICML 2026

TRACE: A Multi-Layer Benchmark for Human AI Controller Coordination Under Drift and Failure

TRACE：漂移与故障下人机控制器协调的多层基准

Joshua Zuniga, Srinivasan Subramanian, Ramya Madhuri Narapureddy, Md Abdullah Al Hafiz Khan

AI总结 该研究构建了名为TRACE的多层基准，向ALFRED轨迹注入受控漂移，经实验验证其可有效识别归因漂移，且重型注意力模型在该符号基准上无优势。

Comments This work was accepted for presentation and publication at the 25th IEEE International Conference on Machine Learning and Applications (ICMLA 2026)