AgentArk：将多智能体智能蒸馏到单个大型语言模型智能体中

Multi-agent systems demonstrate remarkable capabilities in complex tasks, yet their high computational cost and inference latency hinder practical deployment. AgentArk proposes a novel distillation framework to compress the collective intelligence of multi-agent systems into a single Large Language Model (LLM) agent. This framework comprises three core modules: First, it generates high-quality trajectory data through multi-agent collaboration, serving as the teacher signal for distillation. Second, it designs a multi-faceted distillation strategy, encompassing behavior cloning, chain-of-thought distillation, and planning distillation, to comprehensively capture the multi-agent decision-making processes and reasoning abilities. Finally, an adaptive distillation mechanism is introduced to dynamically adjust distillation weights based on the student agent's learning progress. Experimental results show that AgentArk significantly enhances the performance of a single LLM agent across various complex tasks, enabling it to approach or even surpass the original multi-agent system's performance while maintaining high efficiency. AgentArk offers a new avenue for constructing efficient and powerful LLM agents, promising to promote their widespread application in resource-constrained environments.