Multi-Agent Actor-Critic for Decentralized LLM Collaboration

A recent study published on arXiv introduces Multi-Agent Actor-Critic (MAAC) techniques aimed at enhancing decentralized cooperation among large language models (LLMs). The researchers contend that decentralized methods are more feasible than centralized ones, allowing agents to perform inference simultaneously and adaptively. They present two strategies: CoLLM-CC, which utilizes a centralized critic, and CoLLM-DC, featuring decentralized critics. This research tackles the high variance challenges associated with Monte Carlo techniques in existing fine-tuning methods, which necessitate a larger number of samples for effective training. The paper further examines the conditions under which MAAC methods prove advantageous for the collaboration of LLMs.