World models trained with joint-embedding predictive architectures (JEPAs) aim to predict future states directly in latent space, avoiding the cost of reconstructing raw observations. But the arXiv abstract for LeHami warns that unconstrained autoregressive predictors have a failure mode: over long rollouts, they can gradually contract the latent distribution, degrading the model's usefulness.
LeHami addresses this by imposing a symplectic Hamiltonian structure on the latent dynamics. Hamiltonian systems conserve phase-space volume, which in principle prevents the contraction seen in naive predictors. The method is presented as a way to keep latent representations well-spread and stable as predictions extend further into the future.
The paper is announced as new on arXiv, though the abstract is brief and does not include experimental results or comparisons. The core claim is that adding this geometric constraint to JEPA-style world models yields more reliable long-horizon latent prediction.