LeCun Paper Proves AI 'Platonic' Hypothesis

Yann LeCun and Randall Balestriero on Oct. 29 posted a paper to arXiv that claims to turn the Platonic Representation Hypothesis, the idea that AI models trained on different kinds of data drift toward the same internal picture of the world, into a proven theorem. Their experiments suggest a 109-million-parameter text model can line up with vision models better than language models as large as 72 billion parameters.

Key takeaways

• The paper proves that joint-embedding predictive architectures (JEPAs) trained separately on any modalities learn the same representation, up to a single rotation, when their training pairs are two views of the same scene.
• A 109M text JEPA aligned with DINOv2 and CLIP image encoders better than every language model the authors tested, up to 72B parameters, and needed 42 times fewer samples than next-token prediction.
• The result rests on idealized assumptions, and the code is public under an MIT license so others can check it.

What is the Platonic Representation Hypothesis?

The hypothesis was put forward in 2024 by MIT researchers Minyoung Huh, Brian Cheung, Tongzhou Wang and Phillip Isola. They argued that neural networks trained on different objectives, data and modalities are "converging to a shared statistical model of reality in their representation spaces."

"The Platonic Representation Hypothesis is now a Theorem," Balestriero wrote on X when announcing the work. The paper lists Balestriero with Brown University and AMI Labs, and LeCun with New York University and AMI Labs.

How the JEPA proof works

A JEPA learns by predicting the embedding of one view of something from another view, instead of predicting raw pixels or the next word. The authors show that if the positive pairs are "good views," meaning two looks at the same underlying scene, separately trained JEPAs end up with representations that differ only by a rotation.

The authors frame next-token prediction as a JEPA built on "bad views," because its pairs are just pieces of a single document. That, they argue, explains why large language models align with vision models mainly through scale.

The numbers: 109M vs. 72B

The team tested the theory on datasets ranging from 145,000 captions to the texts of 940 million web images, and reports a scaling law for alignment. The language models compared include Qwen2.5 models from 0.5B to 72B, Llama-3-8B and -70B, and Mistral-7B. According to the project's GitHub repository, good views matched next-token prediction's alignment with 21.6 million samples instead of 900 million.

Context and limits

LeCun has long argued that next-token prediction alone will not produce human-level AI, and JEPAs are the architecture he has promoted as an alternative. The paper gives that position a formal footing, but it is a preprint and has not been peer reviewed.

The authors list several caveats. The theorem assumes a finite world and ideal optimization, covers only JEPAs with an identity predictor, and treats each observation as fully determining its scene, which a caption does not do for a photo. Their own image JEPA, trained on 7,000 video clips, is much weaker than DINOv2.

What's next

The authors leave scaling good image views to future work. The prh-proof repository maps every figure and table to its training runs, which gives outside labs a direct way to test whether the 42x efficiency claim holds at larger scale.

FAQ

Does this mean small JEPAs beat large language models?

Only on the narrow measure of alignment with vision encoders and related zero-shot tests. The paper does not claim the 109M model rivals 72B models at writing, reasoning or coding.

Is the result peer reviewed?

No. It is an arXiv preprint posted Oct. 29, 2026, and its theorem depends on idealized assumptions the authors spell out.

References

• Proving and Improving the Platonic Representation Hypothesis with JEPAs — arXiv (Oct. 29, 2026)
• galilai-group/prh-proof — GitHub
• The Platonic Representation Hypothesis — Huh et al., MIT (ICML 2024)