SD-JEPA decomposes JEPA latents into orthogonal progression and content subspaces, improving control benchmarks and yielding a 1D angular coordinate that tracks task progress and localizes semantic events better than prediction error.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Subspace-Decomposed JEPAs: Disentangling Progression and Content in Latent World Models
SD-JEPA decomposes JEPA latents into orthogonal progression and content subspaces, improving control benchmarks and yielding a 1D angular coordinate that tracks task progress and localizes semantic events better than prediction error.