THE ESSENTIALS IN 30 SECONDS

  • The preprint, posted on October 29, proves that JEPAs trained on 'good views' of the same scenes converge to one representation up to a rotation, under stated idealised conditions.
  • On one alignment benchmark, a text JEPA matched next-token prediction with about 42 times fewer training samples drawn from the same corpus.
  • The paper has not been peer reviewed; the authors list limits including small image models, some single-seed results and a heuristic link to language models.

Randall Balestriero and Yann LeCun posted a preprint on arXiv on October 29 that tries to put a well-known observation in AI research on a mathematical footing. The paper, titled "Proving and Improving the Platonic Representation Hypothesis with JEPAs", lists both authors with AMI Labs, alongside Brown University for Balestriero and New York University for LeCun. It has not been peer reviewed.

From hypothesis to theorem

The Platonic representation hypothesis, put forward in an earlier paper, holds that neural networks trained on different data and modalities tend to drift towards similar internal representations of the world. Balestriero and LeCun study joint-embedding predictive architectures (JEPAs), a self-supervised approach in which a model learns by predicting the embedding of one view of something from another.

Their main result says that if separate JEPAs, one per modality, are trained on pairs that are two observations of the same scene, which the authors call good views, their optimal encoders are identical up to a single rotation. A second result bounds how far apart they can drift when the pairs differ. The authors then argue that next-token prediction behaves like a JEPA with bad views, because its pairs are pieces of one document, which would explain why language models only line up with vision models at large scale. They describe that link as heuristic rather than a strict consequence of the theorem.

The proofs rely on simplifying assumptions spelled out in the paper: a finite world, ideal rather than finite-sample training, a fixed identity predictor, and, for the LeJEPA objective used in experiments, uniqueness up to rotation taken as an assumption.

What the experiments show

Tests used Flickr30k with small image and text encoders, captions for up to 940 million web images from Recap-DataComp, and MSR-VTT video clips. Alignment was measured by how often a model's nearest neighbours match those of a frozen DINOv2 vision model. On that metric, a good-view text JEPA reached the same level as next-token prediction on the same corpus with about 21.6 million samples instead of 900 million, roughly 42 times fewer (about 14 times fewer text tokens). A 109-million-parameter text JEPA scored higher than Qwen2.5-72B on the same test. The authors also report that a rotation alone lets separately trained encoders caption images.

Caveats

The 42-fold figure concerns cross-modal alignment, not general language ability. The paper itself notes that most image experiments use a ViT-S trained on 29,000 images, that some results come from a single seed and that training could collapse without gradient clipping. Independent replication has yet to come.