Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster E in Poster Session E: Thursday, August 6, 10:30 am – 12:15 pm, Kimmel Center, Shorin & Rosenthal Rooms

Theory of Mind Improves Multi-Agent Coordination: Accuracy Thresholds and Active Inference

William Beaumaster¹, Paul R. Schrater¹; ¹University of Minnesota

Presenter: William Beaumaster

We investigate how Theory of Mind (ToM)—the ability to infer hidden agent preferences—improves coordination in multi-agent systems. In a 16 × 10 grid-world with 12 agents following one of 10 hidden behavioral rules, a central coordinator must maximize aggregate welfare by placing and removing structures, observing only positions, types, and happiness. We train supervised ToM models (LSTM) achieving 22–95% classification accuracy, finding that temporal observation length is the dominant factor in mental state inference. We train PPO coordinators under three conditions: Blind (no rule knowledge), Oracle (true rules), and ToM-augmented (predicted rule probabilities). Preliminary results show the blind coordinator learns meaningful policies (episode reward improving from -65 to -11). We hypothesize a threshold effect in the accuracy-benefit curve, where ToM accuracy above ∼ 80% yields disproportionate coordination gains, particularly for cascading inter-agent dependency rules.

Topic Area: Methods, Tools, Theory & Neural Coding