A JEPA aligns separate vision and language encoders into a shared representation, enabling tasks like image captioning.
JEPAS PROVE AI MODELS CAN SHARE A SINGLE REPRESENTATION
*The new theorem turns a long‑standing AI conjecture into a provable claim. Randall Balestriero and Yann LeCun show that joint‑embedding predictive architectures align across vision and language far more efficiently than traditional language models.*
On 29 October 2026, researchers Randall Balestriero and Yann LeCun uploaded a paper to arXiv titled “Proving and Improving the Platonic Representation Hypothesis with JEPAs.” The work tackles a core AI question: do models trained on vision, language, or other modalities converge on a single, universal representation of the world? Their answer is a formal theorem that guarantees convergence when training pairs are true observations of the same scene—so‑called good views. The authors also expose why the dominant next‑token prediction paradigm, which relies on fragmented text, lags dramatically behind this approach. By turning a hypothesis into a provable statement, the paper reshapes expectations for multimodal AI development.
What is the Platonic Representation Theorem?
The theorem states that Joint‑Embedding Predictive Architectures (JEPAs) trained separately on any number of modalities converge to the same internal representation, differing only by a single rotation. Convergence occurs when the training pairs are two observations of the same scene—what the authors call “good views.” If the pairs are misaligned, the theorem guarantees that the resulting representations remain bounded in their divergence. This formalizes the Platonic Representation Hypothesis, which previously rested on conjecture alone.
Why next‑token prediction falls short
The authors classify next‑token prediction as a JEPA that uses “bad views,” i.e., fragments of a single document rather than true cross‑modal observations. Because bad views provide weaker alignment signals, language models need far more data to approach the same representation quality. The paper argues that scale alone cannot compensate for the lack of good views, making next‑token prediction an inefficient path to platonic representations.
Empirical validation: scaling law and sample efficiency
Experiments span 145,000 image captions to the texts of 940 million web images. The authors uncover a scaling law that predicts alignment performance within seed noise. Crucially, good views require 42 times fewer training samples than next‑token prediction on the same corpus to reach equivalent alignment. A 109‑million‑parameter text JEPA outperforms every language model evaluated, including models up to 72 billion parameters, and even beats CLIP’s own text encoder.
Real‑world impact: captioning and zero‑shot benchmarks
A single rotation aligns the separately trained encoders into an image captioner that scores CIDEr 0.838 on COCO. On a pair‑free benchmark, the same rotation yields a text JEPA that surpasses all zero‑shot results reported for language models up to 8 billion parameters, achieving 64 % top‑5 accuracy on ImageNet‑100 versus a maximum of 56 % from those models. The results prove that positive views, not sheer scale, drive the construction of shared representations.
Frequently asked questions
What does the Platonic Representation Theorem claim?
It claims that JEPAs trained on any number of modalities learn the same representation up to one rotation when their positive pairs are two observations of the same scene, called good views.
How much more sample‑efficient are good views compared to next‑token prediction?
Good views need 42 times fewer training samples than next‑token prediction on the same corpus to reach the same alignment.
Did the text JEPA outperform larger language models?
Yes, a 109‑million‑parameter text JEPA aligned better than every language model evaluated up to 72 billion parameters and outperformed CLIP’s text encoder.
If the field adopts JEPAs with good views as the standard, future AI systems could achieve cross‑modal understanding with a fraction of the data and compute currently required. The theorem not only validates a long‑standing conjecture but also offers a clear engineering lever—positive views—to accelerate the race toward truly unified intelligence.
Sources: arXiv:2610.39302