LeWAM: End-to-End World and Action Modeling with JEPAs.

arXiv:2610.12818v1 Announce Type: cross Abstract: Joint-embedding predictive architectures (JEPAs) learn compact latent world models without pixel reconstruction. Existing JEPA pipelines typically train visual representations and action-conditioned dynamics be