LeCun Team Releases H-JEPA, an End-to-End Hierarchical World Model

LeCun's team has released H-JEPA (arXiv:2610.06805), the first end-to-end trained hierarchical world model, designed for long-horizon visual planning. On the Visual AntMaze task, planning success rate improved from 18% to 73%. Author Randall Balestriero added that the benefits of hierarchy have been validated across multiple experiments and environments, from simulation to the real world.

Confirmed

H-JEPA was proposed by Yann LeCun, Randall Balestriero, Kevin Ghass, and others; it is the first hierarchically organized JEPA world model learned end-to-end, for long-horizon visual planning.
Core mechanism: it trains a set of action-conditioned JEPA hierarchies where each higher level predicts further into the future within its own latent space; planning proceeds top-down, with the highest level planning toward the goal, its predictions becoming subgoals for the level below.
The authors note that semantic abstraction brings temporal compression: forcing higher levels to discard fast-changing, hard-to-predict details means higher-level representations naturally evolve on slower timescales.
Training: end-to-end training of hierarchical JEPA requires preventing dimensional collapse; the team reused SIGReg to keep latent spaces well-behaved at all levels, combined with LeWM—no complex heuristics needed, with stable training that is easy to debug.
Results: Visual AntMaze success rate rose from 18% to 73%.

Why it matters

H-JEPA validates LeCun's long-advocated world-model and abstraction-hierarchy approach: hierarchical abstraction naturally yields multi-timescale representations, achieved end-to-end rather than through hand-designed hierarchical planning.
The large success-rate jump and sim-to-real validation offer a new path for long-horizon planning, a core challenge in robotics and embodied AI.

2026-10-06 ~ 2026-10-06 · 5 related posts

Primary sources

LeCun's H-JEPA: Hierarchical world model lifts planning success from 18% to 73% — randall_balestr · 2026-10-06 22:34 (UTC+08:00)
H-JEPA: First End-to-End Learned Hierarchical World Model for Long-Horizon Visual Planning — randall_balestr · 2026-10-06 22:34 (UTC+08:00)

LeCun's AMI Unveils H-JEPA: Hierarchical World Models Lift Visual AntMaze From 18% to 73% — arankomatsuzaki · 2026-10-06
[source] H-JEPA: First End-to-End Learned Hierarchical World Model for Long-Horizon Visual Planning — randall_balestr · 2026-10-06
H-JEPA Follow-Up: SIGReg Keeps Latent Spaces Well-Conditioned Across All Levels — randall_balestr · 2026-10-06
H-JEPA author: hierarchical benefits confirmed from simulation to real world, per LeCun's long-held thesis — randall_balestr · 2026-10-06

1 near-duplicate retellings: randall_balestr