Training Object Permanence in World Models
A new benchmark tests whether video world models can track objects that disappear, persist, and obey solidity.
The paper introduces WROP, a dataset of 150 hand-designed object-permanence tasks across six cognitive categories. Its Blender generators vary lighting, speed, camera angle, and other nuisance details while preserving the reasoning setup, producing a 1.5M-sample training corpus and a 300-question exam. The authors evaluate 14 video models and report that their 16B model, PWM-WROP, ranks first among continuation models and third overall in a blind pairwise Elo study. They also release the data, exam, answers, scores, weights, and their PyTorch training stack. HF Daily Papers' note
The paper introduces WROP, a dataset of 150 hand-designed object-permanence tasks across six cognitive categories. Its Blender generators vary lighting, speed, camera angle, and other nuisance details while preserving the reasoning setup, producing a 1.5M-sample training corpus and a 300-question exam. The authors evaluate 14 video models and report that their 16B model, PWM-WROP, ranks first among continuation models and third overall in a blind pairwise Elo study. They also release the data, exam, answers, scores, weights, and their PyTorch training stack. HF Daily Papers' note
score 5