UC San Diego trains 7B agent with partner-belief reward for multi-agent systems

UC San Diego researchers publish Mental-Models for Multi-Agent Systems, a paper on training LLM agents with a theory-of-mind reward.

They train a model that infers a partner's beliefs, intents and thoughts, including what the partner thinks of the agent. During reinforcement learning, that model scores candidate replies from a 7B agent. The final agent runs without the partner model at deployment, according to the researchers' paper as reported by Rohan Paul on X.