Paper Proves Platonic Representation Hypothesis Using JEPAs

Randall Balestriero submitted a paper to arXiv on October 29, 2026, titled 'Proving and Improving the Platonic Representation Hypothesis with JEPAs.' The work addresses the Platonic Representation Hypothesis, which conjectures that models trained on different modalities converge to the same representation of the world as they scale. The authors proved a Platonic Representation Theorem stating that Joint-Embedding Predictive Architectures trained separately on any number of modalities learn the same representation up to one rotation under specific conditions. This theorem applies when positive pairs are two observations of one scene described as good views.

The theory was validated using data ranging from 145k captions to the texts of 940M web images. Good views require 42 times fewer training samples than next-token prediction on the same corpus to reach the same alignment. A 109M-parameter text JEPA aligns with DINOv2 and CLIP's image encoders better than every language model evaluated up to 72B parameters. It also aligns better than CLIP's text encoder. One rotation of separately trained encoders results in an image captioner achieving a CIDEr score of .838 on COCO. The text JEPA beats every zero-shot accuracy reported with language models up to 8B parameters on the latest pair-free benchmark. The authors conclude that positive views, not scale alone, are the lever to build platonic representations.