As multi-agent architectures and agent-to-agent protocols proliferate, a fundamental question arises: what actually happens when autonomous LLM agents interact at scale? We study this question empirically using data from Moltbook, an AI-agent-only social platform, with 800K posts, 3.5M comments, and 78K agent profiles. We combine lexical metrics (Jaccard specificity), embedding-based semantic similarity, and LLM-as-judge validation to characterize agent interaction quality. Our findings reveal agents produce diverse, well-formed text that creates the surface appearance of active discussion, but the substance is largely absent. Specifically, while most agents ($67.5\%$) vary their output across contexts, $65\%$ of comments share no distinguishing content vocabulary with the post they appear under, and information gain from additional comments decays rapidly. LLM judge based metrics classify the dominant comment types as spam ($28\%$) and off-topic content ($22\%$). Embedding-based semantic analysis confirms that lexically generic comments are also semantically generic. Agents rarely engage in threaded conversation ($5\%$ of comments), defaulting instead to independent top-level responses. We discuss implications for multi-agent interaction design, arguing that coordination mechanisms must be explicitly designed; without them, even large populations of capable agents produce parallel output rather than productive exchange.
Edu-Theater: A Data-Efficient Agent Framework for Scalable Learner Behavior Simulation through Staging Roll-Call
University of Science and Technology of China(中国科学技术大学);State Key Laboratory of Cognitive Intelligence(认知智能国家重点实验室);Southeast University(东南大学);Alibaba Group(阿里巴巴集团);iFLYTEK Co., Ltd.(科大讯飞股份有限公司)
The Oscars of AI Theater: A Survey on Role-Playing with Language Models
Hong Kong University of Science and Technology (Guangzhou)(香港科技大学(广州));Hong Kong University of Science and Technology(香港科技大学);Tencent(腾讯);Singapore Management University(新加坡管理大学)
Open Challenges in Multi-Agent Security: Towards Secure Systems of Interacting AI Agents
Oxford Witt Lab, University of Oxford(牛津Witt实验室,牛津大学);Department of Engineering Science, University of Oxford(牛津大学工程科学系);Association for Computing Machinery (ACM)(计算机协会(ACM));Independent(独立);MATS Research(MATS研究);CyLab Security & Privacy Institute, Carnegie Mellon University(CyLab安全与隐私研究所,卡内基梅隆大学);Qualcomm Inc.(高通公司);Oxford Martin AI Governance Initiative(牛津马丁人工智能治理倡议);Carnegie Mellon University(卡内基梅隆大学);MIT Media Lab(麻省理工媒体实验室);SAP SE(SAP德国分公司);OWASP GenAI Security Project - Agentic Security Initiative(OWASP生成式AI安全项目-代理安全倡议);Contramont Research(Contramont研究);The Alan Turing Institute(艾伦·图灵研究所);Department of Economics, New York University(纽约大学经济系);Zenity(Zenity公司);King’s College London(伦敦国王学院);Arizona State University(亚利桑那州立大学);Torr Vision Group, University of Oxford(托尔视觉组,牛津大学);Deep Cyber Ltd(Deep Cyber有限公司)
M2I: From Factored Marginal Trajectory Prediction to Interactive Prediction
Tsinghua University(清华大学);Massachusetts Institute of Technology(麻省理工学院)
预测道路参与者的未来运动是城市场景中自动驾驶的重要任务。现有的模型擅长预测单个代理的边缘轨迹,但它仍然是一个悬而未决的问题,联合预测场景兼容的轨迹超过多个代理。挑战是由于预测空间随着代理数量的增加而呈指数增长。在这项工作中,我们利用相互作用的代理之间的潜在关系,并将联合预测问题解耦为边际预测问题。我们提出的方法M2I首先将交互代理分类为影响者和反应者对,然后利用边际预测模型和条件预测模型分别预测影响者和反应者的轨迹。来自交互代理的预测根据它们的联合可能性进行组合和选择。实验表明,我们简单但有效的方法在Waymo Open Motion Dataset交互式预测基准上实现了最先进的性能。
DynaWeb: Model-Based Reinforcement Learning of Web Agents
Shanghai Jiao Tong University(上海交通大学);Sichuan University(四川大学);Hong Kong University of Science and Technology(香港科学与技术大学);McGill University(麦吉尔大学);Shanghai AI Lab(上海人工智能实验室);Gradient;University of Toronto(多伦多大学);Mila - Quebec AI Institute(魁北克AI研究所)