H-JEPA: hierarchical JEPA world models for long-horizon visual planning

Daily Radar item, dated 7 October 2026. The source text describes a paper led by Yann LeCun that proposes H-JEPA, an architecture that stacks JEPA world models across multiple time scales. Higher levels predict further into the future, and lower levels generate actions. According to the author, success on Visual AntMaze rose from 18% to a value the available excerpt does not fully show. The summary states that hierarchical latent world models can reduce planning cost by handling low-level movement and high-level goals in separate layers.

To consult and verify, search for the original paper by its exact title and check authorship, date, experimental setup and results tables before citing any figure. The excerpt here is a partial summary and does not replace reading the work. The topic is AI research; no personal data handling is involved.