Structured Machine Theory of Mind from Agent Trajectories

TMLR Paper8660 Authors

Abstract: Predictive models of human behavior trained on large-scale trajectory data optimize for statistical accuracy without representing the mental states that causally generate behavior. Such models support prediction but not principled intervention: they cannot answer how an agent's behavior would change if its beliefs or preferences were different. We introduce Structured Machine Theory of Mind (SMToM), a framework that addresses this limitation by attributing explicit, independently supervised belief and desire representations from observed trajectories within a Belief-Desire-Intention causal structure. The central architectural element is a goal head that consumes only the predicted mental-state channels and a current-trajectory embedding, making counterfactual intervention on beliefs and desires a direct operation. We instantiate SMToM on a controlled pedestrian navigation domain where ground-truth mental states are known by construction, enabling rigorous evaluation of both attribution accuracy and counterfactual validity. The resulting model, BDIBottleneck, outperforms trajectory-only and context-aware baselines on top-1 goal inference across path fractions and held-out agent splits, approaching the approximate upper bound at early-to-mid path reveal. Desire counterfactual experiments confirm that substituting an agent's inferred preferences with a different activity type coherently shifts predicted destinations toward relevant locations. Belief counterfactual experiments confirm that marking a location as unavailable in the agent's belief state reliably reduces its predicted probability as a destination, with effects that are statistically significant on both evaluation splits. Together, these results demonstrate, in a controlled navigation setting, that explicit BDI-structured supervision is a viable foundation for causal behavioral analysis of longitudinal trajectory data.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Erin_J_Talvitie1

Submission Number: 8660