Video world models are designed to predict future frames, but a new arXiv paper argues that they must go beyond generating realistic images and instead produce physically plausible futures. The authors note that current methods typically use semantic, physical, or predictive information to steer the visual generation stream, yet the abstract is cut off while describing what this leaves unaddressed.
To close this gap, the paper proposes a "physical registry"—a mechanism for learning internal predictive states that keep the model grounded in physical plausibility. The approach is positioned as a necessary step for world models that aim to reason about how scenes evolve over time.
The full details of the registry mechanism are not available in the source, as the abstract ends mid-sentence. The paper is available on arXiv under identifier 2610.16365.