超越分工：角色如何影响多智能体系统的反馈与恢复
原标题：Beyond Division of Labor: How Roles Shape Feedback and Recovery in Multi-Agent Systems
AI 摘要
该预印本论文研究多智能体大语言模型系统中角色对反馈与失败恢复的影响。作者在Planner–Executor–Reviewer团队中固定智能体、拓扑、调用预算和评估器，仅改变角色指令，在Omni-MATH和SciBench上测试多个开源权重模型。发现删除角色几乎不降低首轮准确率，但角色结构中的Reviewer常拒绝提交，导致团队获得的外部反馈更少、失败后恢复率更低；强制每轮提交可将恢复率从0.198提升至0.613。结论是评估角色应同时考察失败后的恢复能力，而不仅是分工专业化。
正文节选
Preprint Beyond Division of Labor: How Roles Shape Feedback and Recovery in Multi-Agent Systems Abstract. Roles in multi-agent large language model (LLM) systems allocate decisions as well as work. In a Planner–Executor–Reviewer team, the Reviewer’s approval can determine when the team submits an answer and receives corrective feedback. We study both functions by varying role instructions while holding agents, topology, call budget, and ground-truth evaluator constant, using three open-weight mo