AnalysisAI ModelsOctober 6, 2026

H-JEPA: hierarchical world models for visual planning

Read original source →arxiv.org

H-JEPA trains a hierarchy of action-conditioned JEPAs, each level predicting farther ahead in its own learned latent space, with top-down planning where each level's predictions become subgoals for the level below. On Visual AntMaze, a three-level hierarchy raised success from 18% to 73% with less planner compute.

2 sources

More stories today

Open the live feed