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

Learning Mental Models Under Unreliable Evidence

Jeffrey Qin¹, Wasu Top Piriyakulkij², Zhuangfei Gao³, Farzin Ahmadi³, Kevin Ellis², Marta Kryven³; ¹University of Waterloo, ²Cornell University, ³Dalhousie University

Presenter: Jeffrey Qin

Which computational principles define how people construct and refine mental models of the world? We study this question in the context of a puzzle-solving task, where participants integrate misleading social information and experimentation to discover a latent rule in limited time. We argue that real-world learning is supported by (1) the ability to iteratively refine hypotheses to explain partial observations and (2) explicitly representing uncertainty over the reliability of evidence. We show that these principles can be formalized in a unified framework that integrates program induction with Bayesian particle-based inference. This framework admits two complementary interpretations: (i) as a constraint satisfaction process, in which hypotheses must satisfy evidence-derived constraints, and (ii) as a program synthesis problem, in which hypotheses are symbolic programs evaluated against evidence as test cases. We demonstrate that this framework captures human patterns of evidence selection, and outperforming alternative models and its lesioned variants. Together, our results suggest that human learners construct and revise mental models through approximate generative inference that accounts for reliability of evidence.

Topic Area: Memory, Learning & Knowledge Structures