AI research aims to develop autonomous agents that can adaptively operate in complex social environments. Multi-agent reinforcement learning (MARL) methods, however, face significant challenges in these settings, including high sample complexity, poor generalization to unseen agents, and limited reasoning capabilities.

To address the abovementioned challenges, in a new paper Hypothetical Minds: Scaffolding Theory of Mind for Multi-Agent Tasks with Large Language Models, a Stanford University research team proposes Hypothetical Minds, builds on recent advancements in LLM-based agents designed for multi-agent environments, aiming to enhance adaptability in competitive, cooperative, and mixed-motive scenarios with concealed information.

The core innovation of Hypothetical Minds lies in its Theory of Mind (ToM) module, which generates hypotheses about other agents’ latent states. This module anticipates other agents’ strategies, goals, and capabilities, facilitating effective coordination or counter-strategies. High-level plans generated by the ToM module are then passed to a subgoal module, which sequences embodied action plans. The ToM module continuously evaluates multiple hypotheses until it finds one that adequately explains the observed data.

Specifically, the ToM module processes interaction history to produce a target inventory, which serves as a goal for the subgoal module. This information processing occurs in five steps, involving the generation, evaluation, and refinement of hypotheses regarding opponents’ strategies.

The researchers validated the effectiveness of Hypothetical Minds across various multi-agent environments in the Melting Pot benchmark, which includes competitive, collaborative, and mixed-motive domains, as well as 30 distinct evaluation scenarios. The Hypothetical Minds agent significantly outperformed LLM-based and reinforcement learning baselines in every environment and most evaluation scenarios, demonstrating its superior generalizability.

The paper Hypothetical Minds: Scaffolding Theory of Mind for Multi-Agent Tasks with Large Language Models is on arXiv.

Author: Hecate He | Editor: Chain Zhang