AAAI

MARL-GPT: Foundation Model for Multi-Agent Reinforcement Learning

MARL-GPT：多智能体强化学习的基础模型

Maria Nesterova, Mikhail Kolosov, Anton Andreychuk, Egor Cherepanov, Oleg Bulichev, Alexey Kovalev, Konstantin Yakovlev, Aleksandr Panov, Alexey Skrynnik

机构 * MIRAI \& Innopolis University Moscow Russia ； MIRAI \& Innopolis University 纠错

AI总结 本文提出MARL-GPT，一种基于Transformer的多任务模型，能通过离线强化学习在不同多智能体环境中高效学习，无需任务特定调优，实验证明其在多个挑战性任务中表现优异。

Comments Accepted at AAMAS 2026 (AAAI Track)

From Pixels to Personas: Tracking the Evolution of Anime Characters

从像素到人物：动漫角色演变的追踪

Rongze Liu, Jiaxin Pei, Jian Zhu

AI总结 研究通过大规模多模态数据集分析动漫角色演变，结合LLM提取的性格特征与视觉特征，揭示观众群体从儿童向青少年过渡，角色设计呈现萌系化趋势，视觉信号比性格特征更主导观众偏好。

Comments Accepted at the 20th International AAAI Conference on Web and Social Media (ICWSM 2026)

Toward Virtuous Reinforcement Learning: A Critique and Roadmap

迈向道德强化学习：一种批评与路线图

Majid Ghasemi, Mark Crowley

机构 * University of Waterloo（滑铁卢大学） 纠错

AI总结 本文批评了强化学习中常见的伦理模式，提出以美德为中心的替代方法，强调规则导向方法和单一目标强化学习的局限性，并提出通过社会学习、多目标优化、正则化和伦理传统操作化来构建道德强化学习的框架。

Comments Accepted as a workshop paper at Machine Ethics: From Formal Methods to Emergent Machine Ethics workshop at the AAAI 2026 Conference

Enhancing Robustness of Offline Reinforcement Learning Under Data Corruption via Sharpness-Aware Minimization

通过尖锐意识最小化增强对抗数据腐蚀的离线强化学习鲁棒性

Le Xu, Jiayu Chen

AI总结 本文提出通过尖锐意识最小化提升离线强化学习在数据腐蚀下的鲁棒性，通过整合SAM优化器改进IQL和RIQL算法，在D4RL基准测试中显著提升性能。

Comments Accepted as an Oral Presentation at the AAAI 2026 Student Abstract and Poster Program (SAPP)

CharBench: Evaluating the Role of Tokenization in Character-Level Tasks

CharBench：评估字符级任务中分词作用

Omri Uzan, Yuval Pinter

AI总结 CharBench通过大规模字符级任务评估，揭示了分词对字符级任务性能的影响，发现tokenization与正确性弱相关，而词长和字符数更关键，且长token会掩盖字符位置信息。

Comments AAAI-26

VarDrop: Enhancing Training Efficiency by Reducing Variate Redundancy in Periodic Time Series Forecasting

VarDrop：通过减少周期时间序列预测中的变量子冗余来提升训练效率

Junhyeok Kang, Yooju Shin, Jae-Gil Lee

AI总结 VarDrop通过减少周期时间序列预测中的变量子冗余，提升训练效率。该方法利用k-dominant频率哈希进行分组，并通过分层抽样选择代表性token，从而降低计算成本。

Comments Published in AAAI 2025