AAAI

AdaFuse: Accelerating Dynamic Adapter Inference via Token-Level Pre-Gating and Fused Kernel Optimization

AdaFuse: 通过令牌级预门控和融合内核优化加速动态适配器推断

Qiyang Li, Rui Kong, Yuchen Li, Hengyi Cai, Shuaiqiang Wang, Linghe Kong, Guihai Chen, Dawei Yin

AI总结 AdaFuse通过令牌级预门控和融合内核优化，显著降低动态适配器推断延迟，提升LLM的推理效率。

Comments Accepted to AAAI 2026. arXiv admin note: substantial text overlap with arXiv:2405.17741

Attention Gathers, MLPs Compose: A Causal Analysis of an Action-Outcome Circuit in VideoViT

注意力汇聚，MLP组成：视频ViT中一个动作-结果回路的因果分析

Sai V R Chereddy

机构 * Sai V R Chereddy 纠错

AI总结 本文通过因果分析揭示视频ViT中动作-结果回路的机理，发现注意力头负责收集证据，MLP块负责生成成功信号，展示了模型处理人类动作结果的核心计算模式。

Comments Accepted at the AAAI 2026 Workshop on Deployable AI (DAI). Non-archival. Code and custom dataset available upon request

MedEyes: Learning Dynamic Visual Focus for Medical Progressive Diagnosis

MedEyes: 学习动态视觉聚焦以进行医学逐步诊断

Chunzheng Zhu, Yangfang Lin, Shen Chen, Yijun Wang, Jianxin Lin

AI总结 MedEyes通过动态视觉聚焦和双模式探索策略，提升医学逐步诊断的准确性与临床相关性。

Comments AAAI 2026, Medical Chain-of-Thought (CoT), Reinforcement Learning with Verifiable Rewards (RLVR), Multimodal Grounded Reasoning