WROP Dataset Enables Training Object Permanence in Video World Models
September 25, 2026
The WROP infrastructure provides a 1.5M-sample training corpus and a 300-question exam designed to instill human-like cognitive priors in video models. It uses Blender-generated tasks to randomize physical parameters like lighting and camera angles while preserving cognitive structures.
HOW THIS AFFECTS YOU
●
researcherYou can use this structured data to evaluate or train whether video models possess emergent physical reasoning.
●
designerImproved object permanence in world models will lead to more consistent and realistic generative video content.