AI News

The paper investigates whether video generation models, viewed as a class of world models, exhibit object permanence and solidity, two cognitive priors associated with human physical understanding. It introduces WROP, or World Reasoning with Object Permanence, a data infrastructure containing 150 hand-designed tasks across six cognitive categories. Blender generators vary nuisance factors such as speed, lighting, and camera angle while preserving each task's underlying cognitive structure, producing more than 10,000 samples per task. The authors release a 1.5-million-sample training corpus and a 300-question exam. The exam covers 14 video models: three reference-to-video models, seven editing models, and four continuation models, including PWM-WROP, the authors' 16-billion-parameter world model. In a blind pairwise Elo study, PWM-WROP ranks first among continuation models and third overall. The two models ahead of it are reference-to-video models and are statistically tied, according to the paper. The release also includes the dataset, exam, model answers, scores, model weights, and PWM, a native PyTorch training stack for AWS Trainium2.