What it is
Builds a 1.5M-sample benchmark and training corpus for object permanence/solidity in video world models, then fine-tunes a 16B continuation model on it. Instead of probing intuitive physics with small 2D benchmarks or VLM judges, it trains and evaluates video models on 150 hand-designed 3D continuation tasks where occlusion, containment, collision, and support events must be predicted.
Key results
- 1.5M training samples from 150 Blender generators across 6 task families
- PWM-WROP reaches Elo 1679.5, 1st among continuation models
- 3rd overall of 14 models, behind only a tie between 2 reference-to-video systems
Shows that core-cognition-style synthetic data can measurably train stronger physical reasoning in video continuation models.