The Platonic Representation Hypothesis, Now a Theorem — What the JEPA Proof Actually Says

On October 29, Randall Balestriero and Yann LeCun posted a paper to arXiv with a title that dares you to skim it: Proving and Improving the Platonic Representation Hypothesis with JEPAs. The Platonic Representation Hypothesis — coined by Huh et al. for the observation that vision and language models seem to converge toward the same representation of the world as they scale — was, until this paper, an observation. The authors now claim a theorem, and a recipe that reaches the same cross-modal alignment with 42 times fewer training samples than next-token prediction. I read the full text and checked the numbers in it, because the version of this story that fits in a headline is not the version that holds.

What is actually proven. The Platonic Representation Theorem, stated for Joint-Embedding Predictive Architectures: encoders trained separately on any number of modalities learn the same embedding, up to one rotation, whenever their positive pairs — the input pairs the objective declares to be views of each other — are two observations of one scene. The authors call these good views: two photographs of one scene, two captions of one image. Two crops of one photograph are a bad view: they share the camera's viewpoint, the caption's wording — what the paper calls nuisance — and nuisance cannot be rotated away into another modality. That is the whole pivot: alignment is a property of your positive pairs, not of scale. And the fine print matters — the theorem lives in an idealized setting: finite world states, injective renderings, population optima, minimizers unique up to rotation. It is a theorem about spectral JEPAs, not a proof that your favorite chatbot converges to reality.

Next-token prediction, demoted. The second move is the one I find more useful: they recast next-token prediction as a JEPA whose positives are a prefix and its continuation inside one document — a bad view, sharing the document's style. In their own words, this is "a prediction to test rather than a corollary," and they test it on small encoders. The numbers, verified in the full text: on the same corpus, a good-view text JEPA matches the alignment with DINOv2 that next-token prediction reaches at 900M samples using 21.6M samples — the 42× of the headline, and 435× fewer unique images. A 109M-parameter BERT trained on the texts of 940M web images scores .184 on the PRH benchmark's mutual-kNN measure, above Qwen2.5-72B's .154 and CLIP's text encoder's .124 (as the authors state it on the HF paper page). Keep the axis straight: this is an alignment metric with DINOv2, not general capability. The 109M encoder read 116B text tokens; the 72B model it outscores on this metric read 15–18 trillion.

A scaling law, with a shelf life. Alignment follows a fitted law in the fraction of good views and samples seen: after fitting 45 grid cells, it predicted 85 unseen runs before training them at RMSE .0093 — the same .0096 spread you get from rerunning one configuration with a new seed. It breaks at 320M samples for good views, and the authors say so plainly: an empirical summary of one recipe, not a consequence of the theory.

The demo, and the caveats. One linear map bolts frozen DINOv2/DINOv3 encoders to a text-only language model: CIDEr .838 on COCO, with no network in the stack ever trained on image–caption pairs. Pair-free, on the Schnaus et al. benchmark, their 109M text JEPA reaches 64.4 % top-5 on ImageNet-100 against at most 56 % reported for language models up to 8B — though CLIP's text encoder, trained on 400M pairs, still beats it there at 68.7 %. The code is public, MIT, with executable checks for every theorem statement and a reproduce doc for every number, and there is a live demo that captions your photos through one rotation. What it does not have: a good-view JEPA at 7B+ where next-token models are strongest, or good image views at scale — their own image JEPA on video clips is far weaker than DINOv2.

The durable part may not be the theorem. It is the reframing: before you scale, ask what your positive pairs connect. If they connect one document to itself, you are paying 42× for the view someone else gets for free.