RESEARCH · RESEARCH · #1017
WROP dataset and PWM‑WROP 16B world model released to study object permanence
The paper arXiv:2609.28654v1 introduces WROP, a cognitive-science-inspired data infrastructure of 150 hand-designed tasks (six cognitive categories) with Blender generators producing 10,000+ samples per task, and releases a 1.5M-sample training corpus plus a 300-question exam. The authors evaluate 14 video models (3 reference-to-video, 7 edit, 4 continuation), report PWM-WROP (a 16B world model) ranked first among continuation models and third overall in a blind pairwise Elo study, and release data, exam, model answers, scores, model weights, and the PWM native-PyTorch training stack on AWS Trainium2.
KEY POINTS
- The paper arXiv:2609.28654v1 introduces WROP, a cognitive-science-inspired data infrastructure of 150 hand-designed tasks (six cognitive categories) with Blender generators producing 10,000+ samples per task, and releases a 1.5M-sample training corpus plus a 300-question exam.
- The authors evaluate 14 video models (3 reference-to-video, 7 edit, 4 continuation), report PWM-WROP (a 16B world model) ranked first among continuation models and third overall in a blind pairwise Elo study, and release data, exam, model answers, scores, model weights, and the PWM native-PyTorch training stack on AWS Trainium2.
- WROP and PWM‑WROP provide a large, reproducible benchmark, data, and weights specifically targeting object permanence and physical reasoning in world models, enabling systematic study and comparison of emergent physical cognition.
WHY IT MATTERS
WROP and PWM‑WROP provide a large, reproducible benchmark, data, and weights specifically targeting object permanence and physical reasoning in world models, enabling systematic study and comparison of emergent physical cognition.