The paper introduces VR-JEPA, a framework that improves video generation for visual reasoning tasks. It separates state evolution prediction from visual synthesis by using a predictor to generate task-relevant latent trajectories from V-JEPA features. These trajectories then guide a video .
The key innovations are localized contrastive-state learning, which focuses supervision on the most informative states and tokens by comparing successful and generated trajectories, and a skill-routed that allows the predictor to adapt to different reasoning tasks. On the VBVR-Pro-Bench dataset, VR-JEPA achieved a 5.7 percentage point improvement over the VBVR-Wan2.2 baseline.