JEPA · 2026-10-30

Proving and Improving the Platonic Representation Hypothesis with JEPAs

Randall Balestriero, Yann LeCun

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors prove a “Platonic Representation Theorem” for JEPAs: if modalities share the same positive-pair distribution formed from two observations of the same scene (“good views”), their learned representations match up to one rotation. They also show next-token prediction is a “bad-view” JEPA and validate the theory with scaling experiments from 145k captions to texts of 940M web images, reporting that good views are far more sample-efficient than next-token prediction and yield stronger cross-modal alignment.

Why it matters

It turns the Platonic Representation Hypothesis into a theorem with conditions (shared good-view positives and uniqueness up to rotation), and provides an explanation for why language models align with vision only through scale (next-token prediction uses bad-view positives). The paper further validates the theory across datasets and benchmarks, and uses it to build a pair-free image captioner via a single rotation between independently trained encoders.

Method

Prove that in an idealized setting, multiple modalities with shared positive-pair distribution learn the same world embedding up to one rotation each (Platonic Representation Theorem) for whitened JEPAs with spectral gap / uniqueness up to rotation.
Show bounds on misalignment when positive pairs differ (kernel mismatch), and analyze which positives lead encoders to represent the shared “world” vs modality nuisance (good-view vs bad-view JEPAs).
Model next-token prediction as a bad-view JEPA (prefix/continuation positives) and validate via scaling experiments and PRH-protocol alignment metrics on WIT-1024 and additional benchmarks (e.g., COCO, ImageNet-100).

Abstract (from arXiv)

Whether models trained on different modalities converge to the same representation of the world is central to AI. The Platonic Representation Hypothesis (PRH) conjectures that they do as they scale, but without a proof it offers neither a guarantee nor a recipe. We turn the hypothesis into a theorem. First, we prove a \emph{Platonic Representation Theorem}: Joint-Embedding Predictive Architectures (JEPAs) trained separately on any number of modalities learn the same representation up to one rotation whenever their positive pairs are two observations of one scene--\emph{good views}--and their misalignment is bounded otherwise. Second, we show that next-token prediction is a JEPA with \emph{bad views}, parts of a single document, which explains why language models align with vision only through scale. Third, we validate the theory from 145k captions to the texts of 940M web images, and uncover a scaling law of alignment that predicts unseen runs within seed noise. Good views need 42 times fewer training samples than next-token prediction on the same corpus to reach the same alignment, and a 109M-parameter text JEPA aligns with DINOv2 and CLIP's image encoders better than every language model we evaluated, up to 72B parameters, and than CLIP's text encoder: next-token prediction is a remarkably inefficient route to platonic representations. As a showcase, one rotation turns separately trained encoders into an image captioner (CIDEr .838 on COCO). Even that rotation needs no pair: on the latest pair-free benchmark, our text JEPA beats every zero-shot accuracy reported with language models up to 8B parameters, e.g., 64\% against at most 56\% top-5 on ImageNet-100. Positive views, and not scale alone, are thus the lever to build platonic representations.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-29 · Indexed by Pub-AI: 2026-10-30 UTC

preprint