MetaMind: General and Cognitive World Models in Multi-Agent Systems by Meta-Theory of Mind

TL;DR

MetaMind introduces a novel cognitive world model for multi-agent systems that enables agents to infer goals and beliefs of others through self-reflective reasoning, improving collective understanding and task performance without explicit communication.

Contribution

The paper presents MetaMind, a general, self-supervised, meta-theory of mind framework that allows agents to reason about their own and others' beliefs and goals in multi-agent environments.

Findings

MetaMind enables zero-shot reasoning about other agents' goals and beliefs.

Agents with MetaMind outperform baselines in multi-agent task generalization.

MetaMind achieves superior performance in diverse multi-agent simulations.

Abstract

A major challenge for world models in multi-agent systems is to understand interdependent agent dynamics, predict interactive multi-agent trajectories, and plan over long horizons with collective awareness, without centralized supervision or explicit communication. In this paper, MetaMind, a general and cognitive world model for multi-agent systems that leverages a novel meta-theory of mind (Meta-ToM) framework, is proposed. Through MetaMind, each agent learns not only to predict and plan over its own beliefs, but also to inversely reason goals and beliefs from its own behavior trajectories. This self-reflective, bidirectional inference loop enables each agent to learn a metacognitive ability in a self-supervised manner. Then, MetaMind is shown to generalize the metacognitive ability from first-person to third-person through analogical reasoning. Thus, in multi-agent systems, each agent with MetaMind can actively reason about goals and beliefs of other agents from limited, observable behavior trajectories in a zero-shot manner, and then adapt to emergent collective intention without an explicit communication mechanism. Extended simulation results on diverse multi-agent tasks demonstrate that MetaMind can achieve superior task performance and outperform baselines in few-shot multi-agent generalization.