The research announces ARC-Bench, a protocol for evaluating action ranking in frozen JEPA-style world models. It measures whether candidate actions, scored by distances in a frozen latent space, are correctly ordered. The core finding is that this assumption of rankability fails significantly. Audits of official JEPA-WM checkpoints across navigation and manipulation tasks reveal that the top-scored candidate is frequently suboptimal. This issue persists even when using video-pretrained V-JEPA 1 and V-JEPA 2 encoders at ViT-L/ViT-G scale.
The research identifies a masking mechanism as the cause of this invisibility. Reducing the planner's replanning frequency leads to a collapse in success rates, particularly in domains where frequent replanning was previously effective. Furthermore, analysis of rescued episodes reveals a higher prevalence of severe first-plan ranking failures in the PointMaze diagnostic.
This defect systematically overstates the rankability of frozen latent representations. ARC-Bench provides a measurement for this issue and explains the masking mechanism. The work is relevant for methods that adapt, amortize, or replan around latent-space planners without directly auditing released JEPA-WM action rankability.
Source: https://arxiv.org/abs/2609.05461