WROP
A data infrastructure for evaluating physical reasoning capabilities of world models. It contains 150 manually designed cognitive tasks, 1.5 million training samples, and a 300-question test set, focusing on core physical priors such as object permanence and solidity.
Timeline (last 90 days)
PWM-WROP's training corpus contains 1,500,000 samples from 150 generators, with each generator contributing 10,000 samples.
These samples densely cover six cognitive categories, ensuring that the model sees exhaustive combinations of speed, lighting, and viewpoint changes.
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.
WROP provides a reusable data construction methodology—using a generator matrix rather than crawled data to systematically cover the cognitive dimensions of physical reasoning.
Developers are advised to introduce large-scale, cognitively structured video corpora when training continuation-style models, and to rerun existing embodied benchmarks with such datasets as a new baseline.
An 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.
The research validates that core cognitive capabilities such as object permanence can be explicitly trained into video models through carefully designed data, rather than relying entirely on spontaneous emergence from large-scale data.
WROP constructs 150 hand-designed tasks divided into six cognitive categories, corresponding to classic experimental structures used in human cognition research to test physical intuition.
WROP uses a Blender-based generator to randomize disturbance parameters such as velocity, lighting, and camera angle, enabling the generation of more than 10,000 samples per task.
The WROP team released a training corpus of 1.5 million samples as well as an exam set containing 300 questions.
1 more timeline events
All Facts (11)
WROP provides a reusable data construction methodology—using a generator matrix rather than crawled data to systematically cover the cognitive dimensions of physical reasoning.
50%UnverifiedPWM-WROP's training corpus contains 1,500,000 samples from 150 generators, with each generator contributing 10,000 samples.
50%UnverifiedThese samples densely cover six cognitive categories, ensuring that the model sees exhaustive combinations of speed, lighting, and viewpoint changes.
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%UnverifiedDevelopers are advised to introduce large-scale, cognitively structured video corpora when training continuation-style models, and to rerun existing embodied benchmarks with such datasets as a new baseline.
50%UnverifiedThe research validates that core cognitive capabilities such as object permanence can be explicitly trained into video models through carefully designed data, rather than relying entirely on spontaneous emergence from large-scale data.
50%UnverifiedWROP constructs 150 hand-designed tasks divided into six cognitive categories, corresponding to classic experimental structures used in human cognition research to test physical intuition.
50%UnverifiedWROP uses a Blender-based generator to randomize disturbance parameters such as velocity, lighting, and camera angle, enabling the generation of more than 10,000 samples per task.
50%UnverifiedThe research team evaluated 14 video models on the 300-question exam set, covering three paradigms: reference image-to-video (3 models), video editing (7 models), and video continuation (4 models).
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%UnverifiedThe WROP team released a training corpus of 1.5 million samples as well as an exam set containing 300 questions.
50%