Whether models trained on different modalities converge to a shared world representation is a central AI question.
The Platonic Representation Hypothesis (PRH) posits that such convergence emerges as models scale, yet it lacked a formal proof. We elevate the hypothesis to a theorem and present the Platonic Representation Theorem: as long as positive pairs are two observations of the same scene (i.e., good views) and the misalignment of negatives is bounded, Joint‑Embedding Predictive Architectures (JEPAs) trained separately on any number of modalities learn identical representations up to a single rotation $R$.
We further show that next‑token prediction is a JEPA with bad views—only parts of a single document—explaining why language models align with vision only after massive scaling.
Our validation spans from 145k captioned images to the texts of 940M web images. We uncover a scaling law of alignment: good views need roughly 42× fewer training samples than next‑token prediction on the same corpus to reach the same alignment level. A 109M‑parameter text JEPA aligns with DINOv2 and CLIP image encoders better than every language model we tested (up to 72B parameters) and even outperforms CLIP’s own text encoder.
A single rotation $R$ turns independently trained encoders into an image captioner (CIDEr 0.838 on COCO). Remarkably, that rotation requires no paired data: on the latest pair‑free benchmark, our text JEPA beats every zero‑shot result reported for language models up to 8B parameters, e.g., 64% top‑5 on ImageNet‑100 versus a maximum of 56% for competitors.
Thus, positive views—not sheer scale—are the lever for building platonic representations.
Review