Training Object Permanence in World Models
Researchers release WROP and a 16B world model trained for object permanence, ranking near the top of 14 video models.
The paper introduces WROP, a set of 150 hand-designed cognitive tasks in six categories for studying object permanence in video world models. Blender generators yield more than 10,000 samples per task, and the authors release a 1.5 million-sample corpus plus a 300-question exam. Among 14 evaluated video models, their 16B continuation model PWM-WROP ranks first among continuation models and third overall in a blind pairwise Elo study. They also release data, scores, weights, and PWM, a native PyTorch training stack for AWS Trainium2.
- WROP contains 150 tasks across six cognitive categories
- Authors release 1.5 million samples and a 300-question exam
- Fourteen video models were evaluated, including 16B PWM-WROP
- PWM-WROP ranked first among continuation models and third overall
- Weights and the Trainium2 training stack are released
Full article182 words · extracted from huggingface.co · click to collapse
Object permanence and solidity are hallmarks of human cognitive priors. Recent studies show that video generation models, a paradigmatic class of current world models, have begun to show emerged reasoning abilities, making them ideal candidates for building human-like physical intelligence. Do video models have emerged object permanence in them? If not, could we train them with a core-cognition inspired dataset? We introduce WROP (World Reasoning with Object Permanence), a data infrastructure of 150 hand-designed cognitive science inspired tasks, divided into six cognitive categories. We build Blender generators that randomize speed, lighting, camera angle, and other nuisance parameters while preserving each task's cognitive structure, yielding 10,000+ samples per task. We release a 1.5M-sample training corpus and a 300-question exam. On this exam we evaluate 14 video models: 3 reference-to-video, 7 edit, and 4 continuation, among which PWM-WROP, our 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. We release the data, exam, model answers, scores, weights, and PWM, our native-PyTorch training stack on AWS Trainium2.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.28654