Overgenerality from Inference in Perspective-Taking

Timothy Clausner, Christopher Maxey, Matthew D. Goldberg, Paul Zaidins, Justin Brody, Darsana Josyula, Don Perlis

Abstract: The representational and mental processes that support Theory of Mind need greater clarity to guide computational approaches. Perspective-taking is among the key abilities for Theory of Mind. In human visual perspective-taking, two kinds of judgments develop: Level 1 judges whether something can or cannot be seen, and Level 2 aligns perspectives using communication. To investigate these processes, simulated agents with limited visual range, and decentralized active-logic reasoning, moved, saw discrete spatial locations, and communicated locations visited (Level 2); and inferred the locations within other agents’ visual range (Level 1). Controlled conditions systematically varied team size and how far agents spread out. The results are consistent with an account in which Level 1 overgeneralizes and Level 2 corrects, and corrects more, as agents interact more frequently, in larger teams, or by spreading out less. This account of perspectival processes suggests that overgenerality might have both benefits and costs in Theory of Mind, and serve an important role in computation approaches.