Collaborative Belief Reasoning with LLMs for Efficient Multi-Agent Collaboration

TL;DR

CoBel-World introduces a Collaborative Belief World for LLM-based multi-agent systems to achieve efficient, intent-aware collaboration under partial observability. It combines a Symbolic Belief Representation with a Bayesian Belief Collaboration protocol to ground and update agents’ internal beliefs about the environment and collaborators’ intents, enabling proactive communication only when needed. Across TDW-MAT and C-WAH benchmarks, CoBel-World reduces communication cost by 64-79% and improves task efficiency by 4-28% compared with strong baselines, demonstrating the value of explicit belief modeling for robust, scalable collaboration. The work highlights the importance of ToM-like reasoning and structured belief representations for human-like coordination in open-world embodied settings.

Abstract

Effective real-world multi-agent collaboration requires not only accurate planning but also the ability to reason about collaborators' intents--a crucial capability for avoiding miscoordination and redundant communication under partial observable environments. Due to their strong planning and reasoning capabilities, large language models (LLMs) have emerged as promising autonomous agents for collaborative task solving. However, existing collaboration frameworks for LLMs overlook their reasoning potential for dynamic intent inference, and thus produce inconsistent plans and redundant communication, reducing collaboration efficiency. To bridge this gap, we propose CoBel-World, a novel framework that equips LLM agents with a Collaborative Belief World--an internal representation jointly modeling the physical environment and collaborators' mental states. CoBel-World enables agents to parse external open-world knowledge into structured beliefs via a symbolic belief representation module, and perform zero-shot Bayesian-style belief updates through LLM reasoning. This allows agents to proactively detect potential miscoordination (e.g., conflicting plans) and communicate adaptively. Evaluated on challenging embodied benchmarks (i.e., TDW-MAT and C-WAH), CoBel-World significantly reduces communication costs by 64-79% and improves task completion efficiency by 4-28% compared to the strongest baseline. Our results show that explicit, intent-aware belief modeling is essential for efficient and human-like collaboration in LLM-based multi-agent systems.

Figures (16)

• Figure 1: Comparison of existing communication protocols for multi-agent collaboration with our work. From left to right: (a) CoELA Zhang2023BuildingCE: Fixed templates for step-by-step message generation and planning. (b) CaPo Liu2024CaPoCP: Event-driven multi-round discussion. (c) CoBel-World (ours): Belief modeling and adaptive collaboration. Our method enables consistent planning and effective communication.
• Figure 2: Overview of CoBel-World. Cobel-World comprises two key components: (1) Symbolic belief representation: All agents are organized in a collaborative reasoning process to analyze the requirements of the task and summarize the rules in a structured format. The resulting consensus set of belief rules forms the collaborative belief world. (2) Bayesian belief collaboration: After the belief world is constructed, each agent updates it through belief update and belief prediction, both of which are facilitated by LLM reasoning. Adaptive collaborative decisions will be made based on the beliefs.
• Figure 3: Examples of the transformation from unstructured observations to structured beliefs.
• Figure 4: Illustration of the advantages of CoBel-World in terms of planning consistency and communication efficiency on C-WAH benchmark. All methods are powered by GPT-4o. The left part illustrates CoBel-World’s superior planning consistency over CoELA, while the right panel highlights its reduced communication cost compared to CaPo.
• Figure 5: Alice's belief rules construction prompt