Yann LeCun has long argued that language models alone won’t teach AI to understand physics or predict the consequences of its own actions. His bet is on world models — systems that learn how environments behave and support reasoning and planning. That bet now has a concrete result worth studying, even if you never touch a robot.
The single-representation bottleneck
LeCun’s JEPA (Joint-Embedding Predictive Architecture) skips pixel-by-pixel generation. Instead, it converts observations into latent representations and learns to predict future environment states in that abstract space. The problem, as VentureBeat’s coverage relayed by TechOrange explains, shows up on long-horizon tasks: a single-layer model has to simulate many fine-grained steps, which inflates compute and lets prediction errors accumulate.
The maze example makes the tradeoff concrete. A quadruped deciding its route only needs its position, obstacles, and the exit. But controlling each step requires knowing leg joint poses. Force both jobs through one shared representation, and the model has to juggle long-term goals and low-level motion at the same time.
How the hierarchy works
In late 2025, LeCun left Meta to found AMI Labs, which completed a $1.03 billion seed round in March. Now, with academic collaborators, the team has proposed H-JEPA (Hierarchical JEPA). Instead of one prediction level, it learns environment representations at multiple time scales.
According to the H-JEPA research site, planning runs coarse to fine: the top level plans toward the final goal and proposes intermediate subgoals; each lower level refines them, until the bottom level produces actual motor commands. Each layer predicts in its own representation space, and lower layers use the upper layer’s space to check whether their rollouts approach the assigned subgoal. After executing some actions, the model replans from fresh observations.
One design choice stands out. The researchers set the time intervals per level in advance, but they did not hand-specify what each level should retain. End-to-end training worked that out. In the maze experiments, higher-level representations kept position information while discarding fast-changing details like leg poses — exactly the abstraction split that single-layer models can’t make.
What the numbers actually show
The team benchmarked H-JEPA in four simulated navigation and manipulation environments against LeWM (a single-layer JEPA) and HWM (another hierarchical world model). The headline result comes from Visual AntMaze: the single-layer model hit 18% task success, while a three-layer H-JEPA reached 73% — roughly 4x — using less planning compute. Ablation studies suggest the gains come not just from splitting tasks across time scales, but from the top level learning representations better suited to goal planning.
The team also ran offline evaluation on DROID, a dataset of real robot manipulation videos. With an inverse dynamics training signal added, the hierarchical approach produced arm trajectories closer to expert demonstrations, again with lower planning compute.
Limits before you get excited
The main results are from simulation plus offline evaluation of real data — not physical robots autonomously completing long-horizon tasks. The paper also found that adding more hierarchy levels doesn’t always improve performance.
For builders, the transferable idea is the abstraction split: when one representation must serve both strategic and operational decisions, consider letting each layer of your system learn its own state space rather than forcing everything through a shared one. If you’ve followed the earlier discussion of hierarchical planning agents, this is the representation-learning counterpart to that design problem.
Whether H-JEPA holds up on real hardware is the open question — but as a template for structured planning under limited compute, it’s one of the clearer data points the world-model camp has produced.