UC San Diego researchers train 7B AI agent to estimate a partner's beliefs and goals
In brief
- UC San Diego researchers trained a 7B agent to estimate a conversation partner's beliefs and wants.
- arXiv:2610.12453 was posted October 8 and accepted at NeurIPS 2026, per Crypto Briefing.
- Authors report roughly 11.5% relative gain on Sotopia and 26.7 points on BigToM.
- The distilled policy runs alone at test time, requiring no extra models.
- Results are the authors' own and have not been independently replicated.
How the method works
The paper, arXiv:2610.12453, deals with what researchers call theory of mind. It covers two layers. First-order theory of mind covers what a partner believes or wants, while second-order covers what that partner believes about the agent itself.
The agent keeps an explicit internal representation of its partner's hidden mental state (the authors call it an amortized recursive latent state). A belief-conditioned reward model is trained jointly with that representation, and the resulting policy is then distilled so it runs on its own at test time.
That last step is the part worth watching.
Many approaches to reasoning about other agents rely on auxiliary systems running alongside the main model at inference time, according to Crypto Briefing. Per the paper, this one doesn't need extra models or complicated background state calculations once training's done.
The models and the numbers
The main agent is built on Qwen2.5-7B-Instruct, a 7 billion parameter model. It's not a huge model by today's standards. The team also ran experiments on smaller 3B variants and on other model families, including LLaMA and Mistral.
Testing covered two language-only benchmarks (Sotopia and BigToM) plus the multimodal MMRole benchmark. On Sotopia, the authors report the Qwen2.5-7B agent posted an average relative gain of approximately 11.5% across partner settings, measured against a baseline. On BigToM, a belief-tracking test, they report an average gain of 26.7 percentage points on paired true-belief and false-belief accuracy.
Where were the results strongest? According to Crypto Briefing, the headline numbers came from the language benchmarks rather than MMRole, so the multimodal case looks less settled.
What hasn't been verified
The paper was accepted at NeurIPS 2026, according to Crypto Briefing's report on it. That isn't independent confirmation of the findings, and nobody's replicated them yet. Every figure above is the authors' own.
The authors have released their code publicly.
Nothing here amounts to an agent reading minds. The system estimates a partner's beliefs and goals from the conversation, then uses that estimate to decide what to do next. For anyone building multi-agent apps on the web, the claim that a 7B model can do this with no extra runtime machinery is the part to test, and the public code makes that possible.
Frequently asked questions
What is theory of mind in AI agents?
In this paper, 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 UC San Diego approach targets both.
How does the UC San Diego agent model its partner?
The agent keeps an explicit internal representation of its partner's hidden mental state, which the authors call an amortized recursive latent state. A belief-conditioned reward model is trained jointly with it. The policy is then distilled so it runs alone at test time, with no auxiliary models.
What benchmark results did the researchers report?
According to the paper as reported by Crypto Briefing, the Qwen2.5-7B agent posted an average relative gain of about 11.5% on Sotopia and an average gain of 26.7 percentage points on BigToM's paired true-belief and false-belief accuracy. These are the authors' own results and haven't been independently replicated.
