Representation Hypothesis with JEPAs

A paper by Yann LeCun and coauthors argues that text and image encoders align when training treats different descriptions of the same scene as positives, rather than corruptions of one caption. A 109M-parameter JEPA text model scores .184 mutual 10-NN alignment with DINOv2 ViT-B, above Qwen2.5-72B's .154, and reaches next-token-level alignment with 42 times fewer samples. The theory is idealized, one seed scored .121, and the captions partly come from LLaVA, which uses a CLIP encoder. Results are reported by the authors alone.