MoRe：单智能体多角色混合实现多视角协同
原标题：One Model, Many Minds: Unlocking Multi-Agent Synergy in a Single Agent via Mixture of Roles
AI 摘要
arXiv 论文提出 Mixture of Roles (MoRe) 方法，通过可学习码本和查询感知路由器，将多个角色组合成单一 steering vector，在单智能体单次推理中实现多视角专业化。实验显示 MoRe 平均比单智能体基线高 2.2%，性能与多智能体系统相当，但显著降低 token 成本。该方法通过三阶段 SFT 和 GRPO 训练，且保持骨干 LLM 冻结。
正文节选
One Model, Many Minds: Unlocking Multi-Agent Synergy in a Single Agent via Mixture of Roles Abstract Specializing Large Language Models (LLMs) toward distinct abilities underpins successes ranging from personalized assistants to multi-agent systems (MAS). Single-agent paradigms rely on pre-defined personas or steering vectors to induce specialization, yet they impose a single fixed specialization that fails to adapt to diverse queries. Conversely, MAS achieves dynamic multi-perspective problem s