Proving and Improving the Platonic Representation Hypothesis with JEPAs

Researchers led by Randall Balestriero have turned the Platonic Representation Hypothesis into a formal theorem, proving that Joint-Embedding Predictive Architectures (JEPAs) trained separately on different modalities converge to the same world representation—up to a single rotation—when trained on "good views" (two observations of one scene) with bounded misalignment. The paper, posted to arXiv (2610.39302), frames next-token prediction as a JEPA with "bad views," explaining why language models align with vision only through scale, and validates the theory on data from 145k captions to texts of 940M web images, uncovering a scaling law of alignment that predicts unseen runs within seed noise. The efficiency gains are striking: good views require 42 times fewer training samples than next-token prediction on the same corpus, and a 109M-parameter text JEPA aligns with DINOv2 and CLIP image encoders better than every language model evaluated up to 72B parameters, including CLIP's text encoder. As a demonstration, a single rotation converts separately trained encoders into an image captioner scoring CIDEr 0.838 on COCO, while on a pair-free benchmark the text JEPA hits 64% top-5 accuracy on ImageNet-100 versus at most 56% for zero-shot language models up to 8B parameters. The finding matters because it suggests positive views—not scale alone—are the lever for building unified multimodal representations, potentially reshaping how labs allocate training compute across modalities. What to watch: whether major AI labs adopt view-construction strategies over brute-force scaling, and follow-up work extending the theorem beyond the bounded-misalignment conditions.