AAAI

Masked Autoencoders Learn Perception-Relevant Representations from Resting State Neural Data

掩码自动编码器从静息态神经数据中学习与感知相关的表示

Aleksandr Kovalev, Antonio Lozano, Fabrizio Grani, Cristina Soto Sanchez, Leili Soo, Rocío López-Peco, Adrian Villamarin-Ortiz, Roberto Morollón Ruiz, María del Mar Ayuso Arroyave, Alfonso Rodil, Eduardo Fernández

AI总结 研究利用自监督学习，通过在盲人参与者V1区的自发神经活动上预训练掩码自动编码器，测试能否改善感知解码。结果显示该方法有效，表明自发皮层活动有价值，无监督预训练是改善神经解码的好策略。

Comments 6 papers, 4 figures. NeuroAI Workshop, AAAI 2026

Journal ref Proceedings of the First Workshop on NeuroAI Multimodal Intelligence @ AAAI 2026, PMLR 308:93-98, 2026

Separating Clicks from Baits: Using Large Language Models to Detect Misleading YouTube Thumbnails

从诱饵中分离点击：使用大语言模型检测误导性的YouTube缩略图

Wajiha Naveed, Muhammad Muneeb Pervez, Zaeem Mohtashim Khan, Zafar Ayyub Qazi, Zartash Afzal Uzmi

AI总结 研究针对YouTube等平台误导性缩略图问题，提出用大语言模型的多模态检测管道，构建数据集并评估多个模型，发现Claude 3.5 Sonnet性能突出，通过失败分析为视频平台发展提供方向。

Comments Accepted to the 21st International AAAI Conference on Web and Social Media (ICWSM 2027)

Journal ref Proceedings of the International AAAI Conference on Web and Social Media, 2027

Auditing Alignment Controllability in LLMs via Political Axes

通过政治轴审计大语言模型中的对齐可控性

Bartol Bućan, Nikola Sočec, Sarah Isufi, Morena Granić, Luka Hobor, Agneza Krajna, Mihael Kovac, Mario Brcic

AI总结 研究通过对7个大语言模型进行基于提示可控性的压力测试，探讨其在政治轴上的对齐可控性，发现上下文框架起主要作用，模型移动方式有别，强调政治坐标审计需考虑分散性等可控性因素，并发布相关数据和代码。

Comments 17 pages, 6 figures, 4 tables. Accepted at AIES 2026 (AAAI/ACM Conference on AI, Ethics, and Society). This version includes the supplementary appendix. Code and data: https://github.com/mbrcic/llm-political-steerability (Zenodo DOI 10.5281/zenodo.21489805)

Distribution-Specific Curvature Control with Finite-Sample Guarantees for Open-Weight Safety

具有有限样本保证的开放权重安全的特定分布曲率控制

Domenic Rosati, Ali Dadsetan, Hong Huang, Xijie Zeng, Hassan Chowdhry, Subhabrata Majumdar, Hassan Sajjad, Frank Rudzicz

AI总结 研究如何在开放权重模型中防止有害微调并保留良性适应性，提出HarmAlign方法，通过沿估计子空间变形等操作，推导有限样本界，实现条件收敛率控制，在多种设置下有效阻止攻击且保护良性任务。

Comments Under submission AAAI 2027