Fast-LeWM: World Model for Parallel Prediction

jiqizhixin · x · 2026-07-15

Researchers from Xi'an Jiaotong University propose Fast-LeWM, a faster latent-space world model for visual world modeling and planning.

The core idea: encode action prefixes and predict cumulative future states in parallel, rather than step-by-step as in traditional autoregressive rollouts. This reduces both cumulative error and planning time.

In multiple tasks, Fast-LeWM achieves higher average success rates than existing LeWorldModel, while maintaining good accuracy over long time horizons. The post includes links to the paper, project page, code, and institutional report.

Original post →

More from Research

Research channel →