Researchers at the University of California, San Diego have trained a relatively small AI agent to guess what its conversation partner believes and wants. The agent can then act on those guesses without help from any outside model.
The work arrives in a paper titled “Mental-Models for Multi-Agent Systems” (arXiv:2610.12453), posted on October 8, 2026. It has been accepted at NeurIPS 2026. The authors are Hanan Gani, Lulu Shao, and Manmohan Chandraker, all affiliated with UC San Diego.
What the agent actually learns
The core idea borrows from psychology. “Theory of mind” is the ability to recognize that other people hold beliefs, goals, and knowledge that differ from your own.
The paper builds in both layers. First-order theory of mind covers what a partner believes or wants. Second-order theory of mind covers what that partner believes about the agent itself.
The researchers call their approach “mental-model-enabled agents.” The agent maintains an explicit internal representation of a partner’s hidden mental state, stored as what the authors describe as an amortized recursive latent state.
Unpacked, “amortized” means the expensive reasoning work is folded into training. The agent learns a reusable shortcut for estimating someone’s mental state, rather than computing it laboriously from scratch every turn.
Advertisement
The second piece is a belief-conditioned reward model, trained jointly with that mental-state representation. During training, it scores the agent’s behavior based partly on how well it accounts for what the partner believes. That reward signal then shapes the agent’s policy, the rules it uses to choose actions.
According to the research, the policy is distilled so it runs independently at test time. Once training ends, the scaffolding disappears. No extra models ride along, and there are no complicated state calculations happening in the background.
The benchmark numbers
The team built its main agent on Qwen2.5-7B-Instruct, a 7 billion parameter model. It also ran experiments on smaller 3B variants and on other model families, including LLaMA and Mistral.
Testing covered language-only benchmarks, Sotopia and BigToM, plus the multimodal MMRole benchmark, which brings in more than just text.
On Sotopia, which evaluates social interaction skills, the Qwen2.5-7B agent posted an average relative gain of approximately 11.5% across partner settings. That figure measures improvement relative to a baseline, not an absolute score.
The bigger jump came on BigToM, a test of belief tracking. There, the approach delivered an average gain of 26.7 percentage points on paired true-belief and false-belief accuracy.
The authors have released their code publicly, so other researchers can probe the results directly rather than taking the paper’s word for it.
Why smaller agents matter here
This paper tries to give a modest 7B model a specific social skill through better training structure, not sheer scale. Smaller models are cheaper to run and easier to deploy on limited hardware.
Many approaches to reasoning about other agents rely on auxiliary systems running alongside the main model at inference time. Baking the capability into the policy itself sidesteps that overhead.
What this means
The most obvious stakeholders are developers building multi-agent systems. An agent that can estimate what its counterpart believes, including mistaken beliefs, is better positioned to coordinate and can correct misunderstandings. Those are exactly the social interaction skills Sotopia is designed to measure.
The MMRole results suggest the approach is meant to extend beyond pure text, though the strongest headline numbers in the research come from the language benchmarks.
A relative improvement of approximately 11.5% on Sotopia is meaningful, but it is a step forward on a specific test, not a solved problem.
Because the experiments span Qwen, LLaMA, and Mistral families, the approach may not be tied to a single architecture.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.