Object permanence and solidity are considered hallmarks of human cognitive priors, and a new paper asks whether today's video generation models — a leading class of world models — have acquired these abilities. The authors introduce WROP (World Reasoning with Object Permanence), a data infrastructure built around 150 hand-designed tasks inspired by cognitive science and organized into six cognitive categories.
To scale the data, the team built Blender-based generators that randomize nuisance parameters such as speed, lighting, and camera angle while preserving each task's underlying cognitive structure. According to the paper, this yields more than 10,000 samples per task, and the authors release a 1.5M-sample training corpus alongside a 300-question exam.
The exam was used to evaluate 14 video models: three reference-to-video, seven edit, and four continuation models. Among them is PWM-WROP, the authors' own 16B world model. In a blind pairwise Elo study, PWM-WROP ranks first among continuation models and third overall, behind only a statistical tie between two reference-to-video models.
The authors frame object permanence as a foundation of human cognition and describe their contribution as a comprehensive data infrastructure spanning a diverse set of object permanence tasks, each paired with a generator capable of scaling to at least 10,000 varied samples. They report that the infrastructure proved effective for training video models and world models with object permanence.
Alongside the paper, the team says it is releasing the data, exam, model answers, scores, weights, and PWM, a native-PyTorch training stack running on AWS Trainium2.