What's this about?
WROP uses generated cognitive tasks to teach video models that hidden objects continue to exist and solid objects should behave consistently through occlusion and motion.
What exactly is object permanence?
It is the expectation that a ball still exists behind a box even when it is invisible. For a world model, this simple human intuition prevents objects from vanishing or changing identity between frames.
Why does this matter?
author-reported preprint; 25 HF upvotes.
- The released corpus has 1.5 million samples from 150 task templates.
- A 300-question exam compares 14 video models.
- PWM-WROP's high Elo rank is promising, but it is measured on the authors' purpose-built exam.