Today for AI
HOT RADAR
llmHEAT 4.1°

AI CLUSTERED EVENT · 10/10/2026

LeCun's AMI Open-Sources H-JEPA: Hierarchical World Model Achieves 73% Success in Long-Horizon Planning

1 reports archived1 independent sourcesupdated 10/10/2026, 15:42:00
Synthesis & Latest Updates
1 Sources Cross-Validated

Advanced Machine Intelligence (AMI), founded by Yann LeCun, has released and open-sourced H-JEPA, a hierarchical world model for visual planning. By stacking multiple JEPA modules, the architecture allows high-level layers to formulate coarse plans and pass intermediate states as sub-goals to lower layers for detailed execution, effectively balancing long-term objectives with immediate actions. In the Visual AntMaze simulation task, the three-layer H-JEPA achieved a 73% success rate, significantly outperforming the single-layer model's 18%, while requiring less computational overhead.

LATEST/Advanced Machine Intelligence (AMI), founded by Yann LeCun, has released and open-sourced H-JEPA, a hierarchical world model for visual planning. By stacking multiple JEPA modules, the architecture allows high-level layers to formulate coarse plans and pass intermediate states as sub-goals to lower layers for detailed execution, effectively balancing long-term objectives with immediate actions. In the Visual AntMaze simulation task, the three-layer H-JEPA achieved a 73% success rate, significantly outperforming the single-layer model's 18%, while requiring less computational overhead.

HEAT TRENDHourly heat curve

42 fully observed hours
Now
4.1
Peak
10.621:00
24h change
-43%

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

TIMELINECoverage timeline

Total 1 reports · Latest first
  1. 量子位 (微信公众号)T2·86 pts
    • H-JEPA employs a hierarchical structure where high-level layers handle macro-planning and low-level layers refine micro-actions, resolving the conflict between long-term goals and short-term execution.
    • On the Visual AntMaze benchmark, the three-layer H-JEPA achieved a 73% success rate, vastly outperforming the single-layer model's 18%, with higher planning computational efficiency.
    • The project is fully open-source, including code and pre-trained weights, allowing researchers to directly load and reproduce experiments, continuing LeCun's tradition of open research.