Hierarchical JEPA / World Models
Project overview
Exploring hierarchical JEPA-style world models where higher levels are introduced when lower-level predictions leave unresolved structure. The goal is to study whether hierarchy can emerge from residual predictive information rather than being fixed manually.
Core idea
Use latent prediction errors as signals for creating higher abstraction levels.
Motivation
Fixed hierarchy may miss structure; residual structure can indicate where additional abstraction is needed.
Architecture direction
Latent prediction modules, temporal abstraction, top-down constraints, and planning over learned latent states.
Open questions
How stable is emergence of hierarchy across datasets, and how does hierarchy affect long-horizon planning quality?