Recent 3D world models have begun using features extracted from geometry foundation models (GFMs) as predictive world states, leveraging their rich geometric priors. However, as the abstract of a new arXiv paper notes, these features are view-dependent, meaning they entangle the scene representation with the specific viewpoint from which they were computed.

The paper, titled "GeoState: Learning View-Decoupled Geometry States for 3D World Modeling," proposes a method to learn geometry states that are decoupled from the viewing direction. The goal is to preserve the benefits of GFM features while removing the confounding effect of viewpoint, which should make the world model's predictions more consistent across different camera angles.

The abstract is truncated in the source, so specific technical details and experimental results are not available. The contribution, as stated, is the introduction of GeoState as a solution to the view-dependence problem in geometry feature-based world modeling.