S1-MAS：System One 引导计算分工的高效多智能体协作
原标题：Token-Efficient Multi-Agent Collaboration via System One-Guided Computational Division of Labor
AI 摘要
该论文提出 S1-MAS 框架，将多智能体系统中的协调决策与推理任务解耦：由轻量级 System One 控制器（Laya）负责任务选择、角色分配和终止判断，轻量级读取器（Qwen3-0.6B）检索证据，而强 LLM 仅负责开放式推理。在七个基准上对比 AgentVerse、DyLAN 和 SelfOrg，S1-MAS 将 GPT-4o token 消耗降低 44.9%–97.2%，端到端延迟降低 37.8%–93.0%，同时保持更高准确率。
正文节选
Token-Efficient Multi-Agent Collaboration via System One-Guided Computational Division of Labor Abstract Large language model (LLM)-based multi-agent systems (MAS) have become a promising paradigm for complex information-seeking and reasoning tasks by enabling collaborative problem solving among specialized agents. However, existing MAS frameworks tightly couple task reasoning with coordination operations, including task selection, role assignment, message routing, and context management. As int