📄 中文摘要
该研究提出了一种轻量级且无需训练的控制器 REDEREF，用于多智能体大语言模型（LLM）协作，旨在提高递归委派过程中的路由效率。REDEREF 集成了多种技术，包括基于信念的委派，通过汤普森采样优先选择历史上贡献积极的智能体；利用经过校准的 LLM 或程序化评判者进行反思驱动的重新路由；基于证据的选择方法而非输出平均；以及记忆感知的先验知识，以减少冷启动效率低下的问题。该方法在多智能体系统中表现出显著的性能提升，能够有效应对复杂的长时间推理任务。
无训练的自主 AI：多智能体大语言模型系统中的概率控制与协调
📄 English Summary
Training-Free Agentic AI: Probabilistic Control and Coordination in Multi-Agent LLM Systems
This research introduces REDEREF, a lightweight and training-free controller designed for multi-agent large language model (LLM) collaboration, aiming to enhance routing efficiency during recursive delegation. REDEREF integrates several techniques, including belief-guided delegation through Thompson sampling to prioritize agents with historically positive contributions; reflection-driven re-routing using a calibrated LLM or programmatic judge; evidence-based selection instead of output averaging; and memory-aware priors to mitigate cold-start inefficiencies. The proposed approach demonstrates significant performance improvements in multi-agent systems, effectively addressing complex long-horizon reasoning tasks.