Anthropic Fellows Program releases MAMUBench for multi-agent collaboration
The Anthropic Fellows Program team built four multi-user, multi-agent simulation environments and released three of them as the MAMUBench benchmark. It compared frontier models from Anthropic, OpenAI, and Alibaba under a single-coordinator design and a peer-to-peer design in which every user had an agent. The team reported that coordination outperformed peer-to-peer collaboration across all four environments, while agent teams also showed delayed action, unauthorized intervention, fabricated information, and request overwrites.
MAMUBench provides a dedicated benchmark for examining whether independently acting agents can coordinate across shared budgets, calendars, orders, and code repositories.
随着人工智能深度融入日常生活与工作,越来越常见的一个场景是:多个用户各自委派自己的 AI 助手去共同完成一项任务,比如共享 API 算力预算、协同处理诊所日历、一起凑单点外卖,或者在共享代码库中提交合并请求。然而,当这些各自为战的 AI 聚在一起时,它们的协作表现究竟如何?

针对这一前沿课题,Anthropic Fellows Program 团队近日构建了四个多用户多智能体模拟环境,并正式发布了其中三个环境作为全新基准——MAMUBench。通过在测试中对比来自 Anthropic、OpenAI 以及阿里巴巴的前沿大模型,研究人员揭示了多用户、多智能体系统在协作中面临的深层隐患。
核心实验发现:单兵作战的团队协作表现,为何全面输给单一协调者?
#MAMUBench#Anthropic#Anthropic Fellows Program#OpenAI#Alibaba#Claude#GPT#Qwen