Fast LeWorldModel Accelerates Visual Planning with Action-Prefix Prediction

Joint-Embedding Predictive Architectures (JEPAs), including the recent LeWorldModel (LeWM), are promising for reconstruction-free visual world models. However, visual planning with LeWM is computationally intensive due to its reliance on repeated, local one-step latent transition model applications, which also accumulate errors over longer planning horizons. This new research introduces Fast LeWorldModel (Fast-LeWM) to address these limitations. Fast-LeWM replaces the iterative local rollout with an action-prefix prediction mechanism. It encodes action prefixes and predicts future latents in parallel, directly modeling the accumulated effects of actions over multiple horizons. This prefix-level supervision forces the model to learn continuous state evolution, allowing it to evaluate future latents without explicitly simulating intermediate states. Experiments show Fast-LeWM improves success rates and substantially reduces planning time across various tasks, with significantly slower growth in open-loop latent loss.