Training Object Permanence in World Models
QUESTION — Do video generation models possess inherent object permanence, and can training them on a core-cognition inspired dataset instill this physical intelligence?
The authors introduce WROP, a data infrastructure of 150 cognitive-science-inspired tasks yielding a 1.5M-sample training corpus and a 300-question exam to evaluate 14 video models. They build PWM-WROP, a 16B world model trained using a native-PyTorch stack on AWS Trainium2. In a blind Elo evaluation, PWM-WROP ranks first among continuation models and third overall, demonstrating that core-cognition datasets improve physical intelligence in video generators.
WROP comprises 150 hand-designed tasks divided into six cognitive categories.
The training corpus contains 1.5M samples and the exam has 300 questions.
PWM-WROP is a 16B world model ranking first among continuation models.
Hokin · 23 Sept 2026
read the original ↗