Do AI Systems Have World Models? Probing Reasoning, Forecasting, and Generalization

Shiry Ginosar (Toyota Technological Institute at Chicago) https://simons.berkeley.edu/talks/shi... Topics in Intelligence: World Models and Social Reasoning Recent advances in vision and vision-language models have renewed interest in a fundamental question: do these systems acquire internal models of the world? Although such models are trained on images, videos, and other human-produced representations rather than direct experience, they often exhibit behaviors suggestive of knowledge about objects, spaces, and dynamics. Determining what these behaviors reveal about underlying world models, however, remains challenging. In this talk, I will discuss how Kenneth Craik's conception of mental models can provide a useful lens for thinking about this question. Craik argued that internal models enable organisms to reason about alternatives, anticipate future events, and apply past experience to new situations. Building on this perspective, I will examine two recent lines of work that probe these capabilities in modern AI systems: spatial reasoning in language models and motion forecasting in visual models. Together, these examples illustrate both the promise and the limitations of current approaches to studying world models in artificial systems.