OneWM: unified causal training shows world and action modeling benefit each other
An arXiv paper posted on Oct 15, 2026 introduces OneWM, a framework for unified causal training that treats actions both as conditions for world prediction and as targets for policy learning inside a shared video-action architecture, in order to test whether world and action modeling help each other. The authors evaluate policy learning and world modeling on three embodied testbeds — LIBERO, LIBERO-Plus and RoboCasa365 — plus real-world policy experiments. According to the abstract, learning the visual consequences of specified actions improves policy performance and generalization across changing scenes, while action learning improves multi-view world-prediction fidelity in the same architecture, leading the authors to answer both research questions affirmatively.
If these reciprocal benefits hold up, embodied agents could train decision-making and prediction together in one architecture instead of building separate world and policy models.