Position: World Models must live in Parallel Worlds

World models learn spatio-temporal representations that let agents predict future states, interact with environments, navigate, and simulate outcomes. We argue that for generative models to become effective physical-world agents, their world models must support counterfactual simulation: the ability to reason about what-if scenarios and alternative realities. Such parallel-world reasoning can make agents more capable, safe, and creative in novel or out-of-distribution situations, while moving beyond pattern matching toward causal understanding of the world.
OpenReview | Paper