UC San Diego 论文用奖励机制让 7B 智能体学会心理理论
UC San Diego 把读心能力塞进 7B 小模型的训练奖励里，谈判协作类智能体部署时不用再挂额外模型，思路挺巧。
UC San Diego 的论文《Mental-Models for Multi-Agent Systems》提出在 RL 训练奖励中加入对伙伴信念与目标的推断评分。训练时用一个单独模型为候选回复打分，判断回复是否符合推断出的伙伴意图。部署时只保留一个 7B 智能体，不再需要额外模型。相比依赖提示文本或记忆摘要猜测对方目标的常见做法，该方法显式建模伙伴的信念和意图。
A new UC San Diego paper builds theory of mind into a 7B agent through its training reward, so the agent learns to read a partner’s beliefs and goals and then runs on its own with no extra model at deployment.
LLM agents that negotiate, cooperate or role-play usually guess the other side's goals from prompt text or memory summaries. They keep no explicit estimate of the partner's knowledge or plans, so they can misread intent.
They train a model that infers the partner's beliefs, intents and thoughts, including what the partner thinks of the agent. That model scores candidate replies during RL, and the final agent runs alone without it.
For agents that negotiate or collaborate, train with a reward that judges replies against an inferred model of the partner's goals and beliefs.
– arxiv. org/abs/2610.12453
Title: "Mental-Models for Multi-Agent Systems"
