A 109M-parameter text JEPA beats language models up to 72B at aligning with vision

Proving and Improving the Platonic Representation Hypothesis with JEPAs

The Platonic Representation Hypothesis says models trained on different modalities converge on the same world representation as they scale, but it was never proven. This paper turns it into a theorem: JEPAs trained separately on any number of modalities learn the same representation up to one rotation, provided their positive pairs are two views of one scene. Next-token prediction turns out to be a JEPA with bad views, explaining why language models align with vision only through scale. Good views need 42 times fewer samples, and a 109M-parameter text JEPA aligns with DINOv2 and CLIP better than every language model evaluated, up to 72B parameters.