A new preprint on arXiv challenges a common assumption in latent world model research: that a model which predicts the next latent better is necessarily a better model for planning. The authors argue that one-step prediction fidelity is routinely used to select, checkpoint, and ablate JEPA-style world models, yet this metric does not reliably reflect how well the model supports downstream planning.
The paper, titled "One-Step Prediction Fidelity Is a Poor Proxy for Planning Performance in JEPA World Models," reports evidence that this assumption breaks down in practice. While the abstract does not detail the experimental setup, the claim is direct: optimizing for one-step prediction can mislead model development when the ultimate goal is planning.
If confirmed, the finding would push the field to reconsider evaluation criteria for latent world models. Instead of relying on short-horizon prediction accuracy, researchers may need to benchmark planning performance directly or develop proxies that better correlate with it. The preprint is available on arXiv under identifier 2610.19375.