Skip to content
arXiv Game AI· Tianle Chen·· 27 d agoAI score27

BCMAAC: belief-conditioned multi-agent actor-critic for decentralized LLM collaboration

Belief-Conditioned Multi-Agent Actor Critic for Decentralized LLM Collaboration

AI summary

BCMAAC is a belief-conditioned multi-agent actor-critic method for decentralized LLM collaboration. It maintains beliefs about contextual information and uses them to aid training. Experiments on Minecraft, web search, and tool-use tasks show that BCMAAC improves task performance and collaborative behavior over a vanilla actor-critic baseline, with the advantage of belief modeling most pronounced in multi-turn tasks with richer context.

Source: arXiv Game AI · arxiv.org