This page hosts evaluation artifacts for Reasoning-Structured Videos: A Stratified Diagnostic Suite for Compositional Consistency in World Models.
The benchmark tests whether action-conditioned video world models respect three trajectory relations:
Inverse: a path followed by its inverse should return to the initial state.
Loop: a closed path should return to the initial state.
Equivalence: two different paths reaching the same… See the full description on the dataset page:
https://huggingface.co/datasets/VideoWorldmodel/Evaluation.
