about
Symplectic Latent Dynamics for World Models (arxiv.org)
2 points by andsoitis 16 days ago | hide | past | pdf | discuss on HN

In plain words: Instead of guessing states step by step, this world model simulates a physics-style flow with position and momentum that never shrinks the space. On pendulum tasks it predicted velocities better over long rollouts, and kept that space steady where the usual predictor shrank it.

Abstract · LeHami: Symplectic Latent Dynamics for World Models

Joint-embedding predictive architectures (JEPAs) learn world models directly in latent space, but unconstrained autoregressive predictors may contract the latent distribution over long rollouts. We introduce Symplectic Hamiltonian JEPA (LeHami), which replaces the standard predictor with an action-conditioned Hamiltonian flow over a structured latent state (q, p) integrated using an exactly symplectic leapfrog step. This construction preserves phase-space volume at every rollout step and rules out uniform contraction, asymptotically stable fixed points, and merging of distinct latent states, independently of the learned parameters or action sequence. We obtain the momentum-like component from temporal derivatives of the learned position-like representation, providing an explicit second-order state for prediction and control. On conservative and actuated visual pendulum environments, LeHami substantially improves long-horizon velocity prediction over an unconstrained JEPA world model and empirically maintains constant latent volume where the baseline contracts. The learned flow also supports accurate integration at unseen sub-frame step sizes and numerical time reversal. These results show how geometric structure can provide inference-time guarantees that are complementary to representation regularization in latent world models.

Romina Yalovetzky, Randall Balestriero, Yann LeCun
arXiv:2610.15267 · cs.LG, cs.AI, cs.RO · submitted Oct 11, 2026
abstract · pdf · html

add comment on HN