VL-JEPA
Summary
VL-JEPA is a vision-language model that predicts continuous target-text embeddings instead of autoregressively generating text tokens.
Role In The Wiki
VL-JEPA extends JEPA-style representation prediction to general-domain vision-language tasks and selective decoding. It anchors the wiki pattern where language is a readout from a continuous semantic embedding stream, not necessarily the system’s main internal representation.
LeVLJEPA should be treated as the adjacent non-contrastive objective successor, not as the same entity: it keeps cross-modal embedding prediction but replaces contrastive alignment with predictor/stop-gradient asymmetry plus per-modality SIGReg.
Evidence
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in vl-jepa-2025 rather than duplicating verdict rows here.
At the entity level, VL-JEPA extends JEPA-style representation prediction to general-domain vision-language tasks and selective decoding. It anchors the wiki pattern where language is a readout from a continuous semantic embedding stream, not necessarily the system’s main internal representation. This page should stay as the object card; source pages carry slot-level verdicts, evidence, and missing pieces.