Latent world models are designed to predict what happens after an action, but the authors of a new paper argue that accurate prediction alone is not enough. They identify a decision-local prediction gap: among the few futures actually competing for execution, a candidate that appears closer to the goal can still yield a worse real outcome than an available alternative.
To address this, the researchers introduce D-JEPA, a decision-aligned latent world model that learns decision-relevant relations among candidate futures from outcomes that were actually executed. A bounded, permutation-equivariant operator jointly reasons over goal-relative predictive features and ordinal evidence, refining pretrained predictive geometry precisely where action choices matter most.
The method also uses restricted predictor adaptation and a shared ordinal interface to extend this alignment across complementary predictive geometries. D-JEPA then realizes the learned decision structure in JEPA-compatible future representations, allowing deployment through native latent-distance planning.
Evaluations span latent control, manipulation, pretrained action-producing models, physical robots, and autonomous driving. Reported results include 87.89% success on PushT, a 15.04-percentage-point average gain on RoboTwin, and a 17-point gain on physical robot tasks.
The authors position decision-relevant relational structure as a direct bridge between predictive world modeling and effective control. Code, checkpoints, and decision-supervision data are available, alongside a narrated interactive explainer and recorded comparisons at the project site.