Collaborative Belief Reasoning with LLMs for Efficient Multi-Agent Collaboration

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.

💡 Research Summary

The paper introduces CoBel‑World, a novel framework that equips large language model (LLM) agents with a “Collaborative Belief World” – an internal representation that jointly captures the physical environment and the mental states (intents, knowledge, beliefs) of collaborating agents. The authors argue that existing LLM‑based multi‑agent systems rely on fixed communication protocols (step‑by‑step message generation, dense discussion, event‑triggered multi‑round dialogue) and therefore suffer from redundant communication and inconsistent planning, especially under partial observability. CoBel‑World addresses these shortcomings by combining two key components: (1) a Symbolic Belief Representation module and (2) a Bayesian Belief Collaboration protocol.

The Symbolic Belief Representation draws inspiration from planning languages such as PDDL. Beliefs are formalized as tuples of entities, attributes, and predicates, with a special BELIEVE operator that can be nested to express higher‑order mental states (e.g., “Alice BELIEVE Bob BELIEVE apple IN bedroom”). The language supports zero‑order beliefs (direct observations) and first‑order beliefs (beliefs about another agent’s beliefs). To avoid hallucinations, agents collaboratively construct a shared set of belief rules through an iterative propose‑and‑review process, resulting in a consensus belief world that serves as the foundation for later updates.

The Bayesian Belief Collaboration adapts the classic DEC‑POMDP belief‑filter cycle to LLM reasoning. In the update phase, agents receive partial observations consisting of visual perception (o_v) and communicated messages (o_c). They invoke LLMs with specially crafted prompts (“update_zero” and “update_first”) to revise zero‑order and first‑order beliefs respectively. Theory‑of‑Mind (ToM) prompts are employed to ensure that messages are interpreted from the collaborator’s perspective, preventing confusion between private and public information. In the prediction phase, the updated zero‑order belief b₀ₜ, together with the task goal G and progress description P, is fed to the LLM via a “reason” prompt to anticipate future environmental states. The predicted belief (b₀ₜ₊₁) then guides plan generation through a “plan” prompt, yielding an action policy πₜ₊₁ that maximizes task efficiency while minimizing uncertainty.

Experiments are conducted on two challenging embodied benchmarks: TDW‑MAT (a multi‑object transport and arrangement task in a virtual home) and C‑WAH (a collaborative manipulation task with complex physics). Baselines include CoELA, CaPo, RoCo, and other recent LLM‑based collaboration methods. CoBel‑World achieves a 64‑79 % reduction in average communication rounds and improves task completion efficiency by 4‑28 % relative to the strongest baseline. Qualitative analysis shows that the system can proactively detect plan conflicts (e.g., two agents attempting to grasp the same object) and resolve them with minimal intent‑exchange messages (“I will go first”), thereby avoiding costly redundant actions.

The paper acknowledges several limitations: (a) the current belief language handles only up to first‑order beliefs, limiting representation of deeper meta‑beliefs; (b) reliance on full‑scale LLMs incurs significant inference latency, which may be prohibitive for real‑time robotic control; and (c) the Bayesian update is implemented via textual reasoning rather than a mathematically rigorous probability distribution, making quantitative uncertainty estimation coarse. Future directions suggested include extending the language to higher‑order meta‑beliefs, integrating lightweight LLM variants or distillation techniques for real‑time deployment, and hybridizing the textual belief updates with probabilistic graphical models for more precise uncertainty handling.

In summary, CoBel‑World demonstrates that explicit, structured belief modeling combined with zero‑shot LLM reasoning can endow multi‑agent systems with human‑like intent inference and adaptive communication. This advances the state of the art for LLM‑driven collaborative robotics, smart factories, and human‑AI teamwork where partial observability and communication efficiency are critical.