Researchers at UC San Diego trained a 7 billion parameter AI agent to infer its conversation partner’s beliefs and intentions without external model assistance. The approach, called mental-model-enabled agents, incorporates first-order and second-order theory of mind, allowing the agent to model what a partner believes and what a partner believes about the agent itself. Built on Qwen2.5-7B-Instruct, the agent achieved an approximate 11.5% relative gain on Sotopia and a 26.7 percentage point improvement on BigToM for belief tracking. The research, accepted at NeurIPS 2026 and published as Mental-Models for Multi-Agent Systems (arXiv:2610.12453), makes its code publicly available for further validation by the research community.
Source: Read the original article
