PWM-WROP
A 16-billion-parameter world model trained on the WROP dataset, specifically optimized for object permanence and physical reasoning capabilities. It achieves leading performance among video-continuation-style world models.
Timeline (last 90 days)
The study's evaluation used only a single 300-question exam, rather than a broader suite of concrete embodied tasks.
PWM-WROP achieves the best LPIPS of 0.081 on perceptual distance (the second best is 0.105).
PWM-WROP achieves MS-SSIM 0.921 on structural similarity (the second best is 0.877).
PWM-WROP ranks second in CLIP similarity.
The scope of the study leaves open questions: it is unclear whether these gains can persist beyond synthetic WROP generators, and how they perform on longer, uncontrolled video streams has also not been tested.
PWM-WROP is a 16B-parameter world model trained on a customized corpus, and it is the first to let a continuation-style model top the object permanence benchmark.
PWM-WROP's training corpus contains 1,500,000 samples from 150 generators, with each generator contributing 10,000 samples.
In a blind paired Elo study, PWM-WROP ranked first among continuation-style models and third overall, behind only a statistical tie between two reference video generation models.
PWM-WROP ranks first among all continuation-type models and third overall among all 14 models, trailing only two reference image-to-video models.
The team publicly released the 300-question exam set and all model responses, scoring results for each model, PWM-WROP model weights, and the training stack PWM based on native PyTorch running on AWS Trainium2.
2 more timeline events
All Facts (12)
The study's evaluation used only a single 300-question exam, rather than a broader suite of concrete embodied tasks.
50%UnverifiedPWM-WROP achieves the best LPIPS of 0.081 on perceptual distance (the second best is 0.105).
50%UnverifiedPWM-WROP achieves MS-SSIM 0.921 on structural similarity (the second best is 0.877).
50%UnverifiedPWM-WROP ranks second in CLIP similarity.
50%UnverifiedThe scope of the study leaves open questions: it is unclear whether these gains can persist beyond synthetic WROP generators, and how they perform on longer, uncontrolled video streams has also not been tested.
50%UnverifiedPWM-WROP is a 16B-parameter world model trained on a customized corpus, and it is the first to let a continuation-style model top the object permanence benchmark.
50%UnverifiedPWM-WROP's training corpus contains 1,500,000 samples from 150 generators, with each generator contributing 10,000 samples.
50%UnverifiedIn a blind paired Elo study, PWM-WROP ranked first among continuation-style models and third overall, behind only a statistical tie between two reference video generation models.
50%UnverifiedPWM-WROP ranks first among all continuation-type models and third overall among all 14 models, trailing only two reference image-to-video models.
50%UnverifiedVideo continuation is a more difficult task paradigm than reference image-to-video, because the model must extrapolate subsequent physical evolution based on existing clips.
50%UnverifiedThe team publicly released the 300-question exam set and all model responses, scoring results for each model, PWM-WROP model weights, and the training stack PWM based on native PyTorch running on AWS Trainium2.
50%UnverifiedAn arXiv paper (arXiv:2609.28654) proposes a data infrastructure called WROP (World Reasoning with Object Permanence) and trains a 16B-parameter world model PWM-WROP.
50%