---
title: H-JEPA (Hierarchical Joint Embedding Predictive Architecture)
type: vocabulary
url: "https://www.envisioning.com/vocab/h-jepa-hierarchical-joint-embedding-predictive-architecture"
summary: A world model that stacks JEPA predictors across timescales, each level planning in its own latent space and passing subgoals down to finer levels for visual planning.
year: 2026
generality: 0.35
---

# H-JEPA (Hierarchical Joint Embedding Predictive Architecture)

A world model that stacks JEPA predictors across timescales, each level planning in its own latent space and passing subgoals down to finer levels for visual planning.
H-JEPA is a world-model architecture that stacks multiple JEPA (Joint Embedding Predictive Architecture) models across timescales. Each level predicts farther into the future in its own learned latent space, rather than all levels sharing one representation. Wancong Zhang, Basile Terver, Michael Rabbat, Yann LeCun, and Randall Balestriero, working across NYU, AMI Labs, INRIA Paris, and Brown University, introduced it in the October 2026 paper "H-JEPA: End-to-End Learning of Hierarchical World Models for Visual Planning" (arXiv:2610.06805). The paper targets a limit of flat JEPA world models: planning far ahead in a single latent space and timescale forces one representation to carry both fine motor detail and abstract goals, and lets small per-step prediction errors accumulate over many rollout steps.

Each level of H-JEPA encodes observations and predicts action-conditioned latent futures at a coarser time stride than the level below it. In the paper's simulated environments, the three levels span roughly five, ten, and twenty environment steps. All levels train jointly end to end on a prediction loss plus SIGReg, a regularizer added to prevent the latent space from collapsing. Planning runs top-down: the highest level optimizes a coarse action sequence toward the goal in its own latent space, and each level's predicted states become subgoals that the level below refines, until the lowest level outputs primitive actions. The authors describe this as an extension of hierarchical latent planning rather than its origin, building on HWM (Hierarchical World Models), a 2026 method that uses multiple prediction horizons but keeps all levels in one shared latent space.

Across four simulated navigation and manipulation benchmarks, hierarchical planning outperformed a flat JEPA baseline (LeWM) on average. The clearest gain came on Visual AntMaze, where a three-level hierarchy raised success from 18.0 percent to 73.3 percent while using less planner compute, measured in TFLOPs per episode. With added inverse-dynamics supervision, the same recipe extended to real-robot videos from the DROID dataset, where hierarchy improved offline planning fidelity at lower planner compute. The paper reports results only on simulated tasks and an offline robot-video evaluation. It does not include closed-loop deployment on a physical robot.

---
Source: Envisioning — Technology Research Institute (https://www.envisioning.com/vocab/h-jepa-hierarchical-joint-embedding-predictive-architecture)