CMU's WROP trains object permanence into a 16B world model and releases the corpus, weights and a Trainium2 training stack

arXiv / HuggingFace Daily Papers·medium signalXBlueskyLinkedInCopy link

arXiv 2609.28654 (23 Sep, top of HF Daily Papers with 192 upvotes) builds 150 cognitive-science tasks in six categories using Blender generators that randomize speed, lighting and camera. That yields 10,000+ samples per task, a 1.5M-sample training corpus and a 300-question exam. Across 14 video models, the team's 16B PWM-WROP ranks first among continuation models and third overall in blind pairwise Elo. The release includes data, exam, model answers, weights and PWM, a native-PyTorch training stack for AWS Trainium2, which world-model papers rarely ship.