UCSD paper bakes theory of mind into RL reward, letting a 7B agent model partners solo
rohanpaul_ai · x · 2026-10-12
A UC San Diego paper, "Mental-Models for Multi-Agent Systems" (arXiv:2610.12453), introduces mental-model-enabled agents: an inferred model of a partner's beliefs, intents and goals — including second-order beliefs about the agent itself — scores candidate replies during RL via an amortized recursive Theory-of-Mind representation. The trained policy runs standalone at inference without carrying the partner model, yet retains explicit partner modeling. Across negotiation, cooperation, competition and assistance settings in both language and multimodal benchmarks, explicit mental-state modeling consistently improves interaction quality and ToM performance over prompt- and memory-based baselines.
More from coding & agent
- mitsuhiko's weekend 'software factory' with his kids shows how good AI coding models have gotten — mitsuhiko · 2026-10-12
- yes-if: PayPal hackathon project turns payment risk 'no' into the cheapest conditional 'yes, if' — damnGruz · 2026-10-12
- Edit videos with Claude Code: 22-min HyperFrames tutorial goes free — toolstelegraph · 2026-10-12
- Devs now hand agents ambitious tasks and wake up to them solved — tedddyoweh · 2026-10-12
- Indie dev's multilingual launch-video Skill sells at $4.99, logs 2,727 downloads — yihui_indie · 2026-10-12
- After AI Support Cut 33 Agents, Company Wants Human-in-the-Loop for Sales — chenggiskhann · 2026-10-12