Researchers from the University of California, San Diego, developed a method to equip compact AI agents with a theory of mind. The study, titled Mental-Models for Multi-Agent Systems, was accepted for the NeurIPS 2026 conference. The authors are Hanan Gani, Lulu Shao, and Manmohan Chandraker.
The approach enables agents to maintain an internal representation of a partner’s mental state, covering both first-order and second-order beliefs. By integrating a belief-conditioned reward model during training, the researchers distilled the agent's policy so it operates independently at test time without requiring auxiliary systems.
The team primarily utilized the Qwen2.5-7B-Instruct model, while also conducting experiments with 3B variants and other architectures, including LLaMA and Mistral. The method was evaluated using the Sotopia, BigToM, and MMRole benchmarks.
On the Sotopia benchmark, the 7B agent achieved an average relative gain of approximately 11.5%. In tests on BigToM, which measures belief tracking, the model showed an average improvement of 26.7 percentage points in accuracy. The researchers have made their code publicly available.
