Proving and Improving the Platonic Representation Hypothesis with JEPAs
This arXiv paper argues that models trained on different data types can converge on a shared representation, and that it can be proven under specific conditions, not just hoped for as scale grows. Its central finding is that the "good views" used to train joint-embedding predictive architectures, not model size alone, are what drive that alignment.
- The authors prove that JEPAs trained separately on different modalities learn the same representation up to a rotation, provided their positive pairs are two views of the same scene.
- Next-token prediction is framed as a JEPA with "bad views" (parts of one document), which they argue explains why language models align with vision only through scale.
- A 109M-parameter text JEPA aligns with DINOv2 and CLIP image encoders better than every language model tested up to 72B parameters, and needs 42 times fewer samples than next-token prediction to reach the same alignment.
The paper tries to turn the Platonic Representation Hypothesis, the idea that models trained on different data types (text, images) converge on a shared internal picture of the world as they scale, into a provable result. The authors' main theorem says that Joint-Embedding Predictive Architectures (JEPAs) trained separately on any number of modalities end up learning the same representation, up to a single rotation, provided their positive pairs are "good views": two observations of the same scene. If the pairs are misaligned, the result still holds with a bound on how far off the alignment can be. Their second claim is that next-token prediction, the standard way language models are trained, is really a JEPA working with "bad views," meaning different parts of the same document. That framing explains why language models only drift toward visual representations through sheer scale, rather than by design.
The empirical side is where the numbers get interesting. The authors test the theory on everything from 145k captions up to the text attached to 940M web images, and they fit a scaling law for alignment that predicts held-out training runs within seed noise. Their headline comparison: good views reach the same alignment with 42 times fewer training samples than next-token prediction on the same corpus. A 109M-parameter text JEPA aligns with DINOv2 and CLIP's image encoders better than any language model they tried, up to 72B parameters, and better than CLIP's own text encoder. The authors read this as evidence that next-token prediction is a very inefficient route to shared representations. They also demonstrate the practical payoff: a single rotation applied to the separately trained encoders produces an image captioner scoring 0.838 CIDEr on COCO, without any paired caption-image data for that step.
The zero-shot results are the most striking claim, and they deserve some caution. On a pair-free benchmark, the text JEPA reportedly beats every zero-shot accuracy reported for language models up to 8B parameters, reaching 64% top-5 on ImageNet-100 against at most 56% for the language models. The paper is a single v1 submission from late October 2026, with no independent replication yet, and the comparisons depend on the authors' choice of baselines and benchmarks. The core message, that the quality of the views matters more than model size alone, is clearly stated and testable. Whether the proof's assumptions, especially the bounded-misalignment condition for imperfect pairs, hold in messier real data is the question to check before building on it.