LeJEPA
Summary
LeJEPA is a JEPA training objective that combines predictive alignment with Sketched Isotropic Gaussian Regularization. The original LeJEPA source argues for Gaussian embeddings as a collapse-resistant SSL objective; the later identifiability paper narrows the world-model claim to Gaussian latent worlds where the learned state can be recovered up to rotation.
Role In The Wiki
LeJEPA is the central theory-forward source for collapse-resistant JEPA training.
Evidence
Relation To Foundation TSFM Agenda
Use the source-level agenda mappings in lejepa-2025 and lejepa-identifiability-2026 rather than duplicating verdict rows here.
At the entity level, LeJEPA is the central theory-forward object for collapse-resistant and identifiable JEPA training. This page should stay as the object card; source pages carry slot-level verdicts, evidence, assumptions, and missing pieces.