Randall Balestriero and Yann LeCun submitted “Proving and Improving the Platonic Representation Hypothesis with JEPAs” to arXiv on October 29, 2026. The authors say their theorem applies to separately trained JEPA encoders whose positive pairs are two observations of the same scene, with misalignment bounded otherwise, and that the encoders learn equivalent representations up to one rotation. They report experiments ranging from 145,000 captions to text associated with 940 million web images. Their reported results include requiring 42 times fewer training samples than next-token prediction for the same alignment, 64% top-5 accuracy on ImageNet-100 for a 109-million-parameter text JEPA, and a 0.838 CIDEr score on COCO after applying one rotation.