Theory of Mind in Multi-Agent Systems
Theory of Mind (ToM) refers to agents’ ability to model other agents’ beliefs, goals, intentions, and likely actions. In multi-agent systems, ToM enables agents to anticipate collaborators’ behavior and coordinate without explicit communication.
When agents maintain shared belief state representations about each other, they can track goals and actions, leading to smoother coordination and emergent collaborative behaviors. Research shows that LLM agents can exhibit higher-order ToM capabilities, reasoning about what other agents believe about what they believe.
The limitation: ToM in current LLM-based agents remains imperfect. Long-horizon planning and managing hallucinations about other agents’ states remain challenges. Agents may develop incorrect models of collaborators that compound over time.
What makes ToM powerful for MAS: it reduces the need for explicit coordination messages. Agents can infer what others will likely do and adapt accordingly, similar to how experienced human teams can work efficiently with minimal communication.