Hierarchical Task Knowledge Constrains and Simplifies Action Understanding

Abstract

Human social interactions require understanding and predict- ing other people’s behavior. A growing body of work has found that these inferences are structured around an assump- tion that agents act rationally and efficiently in space. While powerful, this view treats action understanding in a vacuum, ignoring that much social inference happens in the context of familiar, hierarchically structured events (e.g.: buying gro- ceries, ordering in a restaurant). We propose that social and world knowledge is critical for efficiently interpreting behavior and test this idea through a simple block-building paradigm, where participants infer an agent’s sub-task (study 1a), next action (study 1b), and higher-level goal (study 1c), from very sparse observations. We compare these inferences against a Bayesian model of goal inference that exploits task structure to interpret agents’ actions. This model fit participant judg- ments with high quantitative accuracy, highlighting how world knowledge may help support social inferences in a rich and powerful way.

Reconciling truthfulness and relevance as epistemic and decision-theoretic utility.

“…Decision-theoretic relevance applies the same theory of mind to everyday discourse, assuming the speaker models the listener as a rational agent. Implicit world knowledge supplies the set of possible actions and associated utilities, while theory of mind supplies the listener's decision process (Davis & Jara-Ettinger, 2022). The speaker's knowledge need not be perfect.…”