Proving and Improving the Platonic Representation Hypothesis with JEPAs

[Submitted on 29 Oct 2026]

View PDF HTML (experimental)

Abstract:Whether models trained on distinct modalities converge to the identical depiction of the earth is chief to AI. The Platonic Representation Hypothesis (PRH) conjectures that they do as they scale, but without a evidence it offers neither a justify nor a recipe. We rotate the assumption into a theorem. First, we demonstrate a \emph{Platonic Representation Theorem}: Joint-Embedding Predictive Architectures (JEPAs) trained separately on any figure of modalities study the identical depiction up to one rotation whenever their affirmative pairs are two observations of one scene--\emph{good views}--and their misalignment is bounded otherwise. Second, we display that next-token prediction is a JEPA alongside \emph{bad views}, parts of a sole document, which explains why tongue models align alongside imagination lone through scale. Third, we validate the theory from 145k captions to the texts of 940M web images, and reveal a scaling law of alignment that predicts unseen runs inside kernel noise. Good views need 42 times small training samples than next-token prediction on the identical corpus to attain the identical alignment, and a 109M-parameter content JEPA aligns alongside DINOv2 and CLIP's depiction encoders improved than all tongue example we evaluated, up to 72B parameters, and than CLIP's content encoder: next-token prediction is a remarkably inefficient path to platonic representations. As a showcase, one rotation turns separately trained encoders into an depiction captioner (CIDEr .838 on COCO). Even that rotation needs no pair: on the latest pair-free benchmark, our content JEPA strikes all zero-shot accuracy reported alongside tongue models up to 8B parameters, e.g., 64\% against at most 56\% top-5 on ImageNet-100. Positive views, and not measure alone, are thus the lever to build platonic representations.