Knowledge-Informed Policy Structuring for Multi-Agent Collaboration Using Language Models

Abstract

This paper proposes a multi-agent cooperative reinforcement learning method guided by large language model strategies. It addresses issues in traditional multi-agent reinforcement learning, such as policy instability and low exploration efficiency. By introducing strategy guidance generated by large language models, the method helps agents converge quickly to optimal cooperative policies. First, a fusion mechanism between the language model and the reinforcement learning framework is established. A dynamic guidance mechanism is designed to adaptively adjust the strength of guidance. This enhances policy stability and increases cooperation success rates across the system. Experiments show that the proposed method outperforms traditional joint policy approaches under various task complexities, agent scales, and guidance strength settings. These results validate the effectiveness of the proposed strategy guidance mechanism.