Balestriero and LeCun prove JEPAs trained on 'good views' learn the same representation up to a rotation
The arXiv preprint's proof holds under idealised assumptions; it reports a 109M-parameter text JEPA aligning with image encoders better than LLMs up to 72B.