[Tiny Paper] Shortcut World Models: Learning to Leap, Not Step
Abstract
Autoregressive world models chain single-step predictions, requiring N forward passes for N steps into the future. We introduce Shortcut World Models, trained to predict environment dynamics across multiple horizons, enabling direct leaping to any learned step-size in a single pass rather than iteratively stepping through intermediate states. Beyond speed, skipping intermediate predictions also improves accuracy: errors compound through state discontinuities in autoregressive rollout, but shortcuts sidestep this accumulation entirely. At inference, adaptive chaining decomposes arbitrary horizons into learned sub-steps, handling step-sizes beyond training while maximizing accuracy with minimal sacrifice in speed. On discontinuous particle dynamics, Shortcut World Models achieve 33–64× fewer forward passes with up to 50% lower error, demonstrating a path toward learned simulators and model-based planning that are both faster and more accurate.