Collaborative Belief Reasoning with LLMs for Efficient Multi-Agent Collaboration

Multi-agent systems powered by large language models struggle in partially observable environments. They miscoordinate by duplicating efforts or communicating redundantly, as they can't infer collaborators' intents reliably.

CoBel-World equips LLM agents with a Collaborative Belief World modeling physical states and mental states. Agents use symbolic beliefs updated via LLM-driven Bayesian inference to detect conflicts and communicate only when necessary.

On TDW-MAT and C-WAH benchmarks, it reduces communication by 64-79% and improves efficiency by 4-28% versus baselines, preserving success rates.

Builders gain scalable multi-agent setups for simulations or orchestration with lower token costs and better uncertainty handling.

Takeaway: Implement belief tracking in agents: dict[agent_id] = {'intent': str, 'confidence': float}; update with LLM on observations. Measure comm reduction on your tasks.

Source: Zhimin Wang et al. — ArXiv cs.AI, May 2026