Video world models need to know how a scene changes even when it is not visible, so that content looks consistent when it reappears in the camera view. A new arXiv paper, "Momentum: Persistent Dynamics for Video World Models" (arXiv:2610.23423), argues that existing approaches are not built for this task. The abstract says current methods depend on observation history, static point clouds, or image-based representations.

According to the abstract, those approaches fall short when content reappears after leaving the camera view. The paper proposes "persistent dynamics" as a way to maintain a model of the scene's evolution beyond the current observation. The title points to momentum as the central idea: scene state should carry forward over time.

The abstract does not include experimental results or architectural details, so the contribution is best understood as a problem framing and a proposed direction. The paper is available on arXiv.