WROP: 150 cognitive tasks train object permanence into video world models

Haotian Zhang · hf · 2026-09-25

Object permanence is a hallmark of human cognitive priors. Researchers introduce WROP (World Reasoning with Object Permanence), a core-cognition-inspired data infrastructure of 150 hand-designed cognitive science tasks across six categories. Blender generators randomize speed, lighting, and camera angle while preserving each task's cognitive structure, yielding 10,000+ samples per task.

They release a 1.5M-sample training corpus and a 300-question exam, then evaluate 14 video models (3 reference-to-video, 7 edit, 4 continuation) including their 16B world model PWM-WROP. In a blind pairwise Elo study, PWM-WROP ranks first among continuation models and third overall, behind a statistical tie of two reference-to-video models. Data, exam, scores, weights, and the native-PyTorch PWM training stack on AWS Trainium2 are all open-sourced.

Original post →

More from Multimodal

Multimodal channel →