Object permanence and solidity are fundamental human priors. Recent work shows that video generation models, a representative class of world models, have begun to exhibit reasoning abilities, making them promising for building human‑like physical intelligence. We introduce WROP (World Reasoning with Object Permanence), a dataset of 150 hand‑crafted tasks inspired by cognitive science, organized into six categories. Blender generators randomize speed, lighting, camera angle and other nuisance factors while preserving each task's cognitive structure, producing over 10,000 samples per task and a total of 1.5 M training instances, together with a 300‑question evaluation exam. We benchmarked 14 video models—3 reference‑to‑video, 7 edit, and 4 continuation—including our 16B world model PWM‑WROP. In a blind pairwise Elo study, PWM‑WROP ranked first among continuation models and third overall, behind a statistical tie of two reference‑to‑video models. We release the dataset, exam, model answers, scores, weights, and the native PyTorch training stack PWM, runnable on AWS Trainium2.
Review