Proving and Improving the Platonic Representation Hypothesis with JEPAs

The paper proves for JEPAs that separately trained modalities with shared positive pairs learn the same representation up to a rotation, and shows empirically that good views align better and more efficiently than bad views or next-token prediction.

The Platonic Representation Hypothesis observes that vision and language models become increasingly similar with scale, but the paper states it remains an observation. For the self-supervised objectives used in practice, especially JEPAs, there is no proof, no prediction of when the hypothesis holds, and no recipe to improve alignment. Scaling models and data is therefore the main lever used so far.

The paper formalizes a world, modalities that observe it, and their positive pairs. It argues that a JEPA with whitened second-order anti-collapse learns the spectral embedding of its positive-pair kernel. When modalities render one world and their positives relate the same world states, their training problems are one problem in different coordinates and their solutions coincide up to a rotation. When positives differ, the paper bounds misalignment by kernel mismatch and the spectral gap. It defines good-view versus bad-view JEPAs, proves that bad views can spend embedding directions on modality nuisance, and shows that next-token prediction has the form of a bad-view JEPA. It then validates the predictions with controlled text, image, and video experiments, a scaling law, a captioner, and pair-free alignment.

• For k=32, the scene explains .804 of the good-view JEPA embedding, .655 of the bad-view JEPA and .608 of next-token prediction; the nuisance bound gives best-rotation errors 12.6, 22.1 and 25.1, while measured errors are 33.5, 39.4 and 40.3. §3.1 Good Views Align Across Modalities
• With a frozen DINOv2 ViT-B/14, the good-view JEPA retrieves the true caption first for 24.6% of test images, about twice the bad-view JEPA .133, next-token .118 and Qwen2.5-0.5B .135; its mutual 10-NN is .204-.257 against .189 and .167 for reference encoders. §3.4 Good-View JEPAs Beat a Pretrained Language Model and Approach Pair-Trained Encoders
• A 109M encoder reaches .109 on WIT after 21.6M samples, the alignment next-token prediction reaches after 900M samples at .111, which is 42 times fewer samples, 14 times fewer text tokens and 435 times fewer unique images; at matched samples the gap is .108 against .071 for the 33M encoder. §4.2 Good Views Need 42 Times Fewer Samples
• The 109M text JEPA trained on texts of 940M web images reaches .184 on WIT with DINOv2 ViT-B, above Qwen2.5-72B .154 and CLIP text encoder .124; with CLIP ViT-H it reaches .230, and on COCO .161. §4.3 A 109M Text JEPA Beats Language Models up to 72B Parameters
• A scaling law fits 45 cells up to 80M samples with RMSE .007 and predicts 85 unseen runs with RMSE .0093, while two seeds of one run differ by .0096; along iso-alignment lines q is proportional to M^-0.30 and unique images enter weakly with gamma .19. §4.4 A Scaling Law for the PRH
• With no pairs, the 109M good-view JEPA reaches FOSCTTM .018 and zero-shot accuracies of 86.0 CIFAR-10, 67.2 CIFAR-100 and 64.4 ImageNet-100; Qwen3-8B rerun reaches .018, 83.4, 61.7 and 47.6, and the JEPA beats it by 5.5 and 16.8 points on CIFAR-100 and ImageNet-100. §5.2 Alignment Without Any Pair

• The theorems are population statements over a finite world with injective renderings and matched marginals; real modalities are lossy, and finite-sample training relies on the network's inductive bias.
• Uniqueness up to rotation is proved for second-order anti-collapse, not for LeJEPA's full-distribution constraint; the paper states that no experiment meets the hypotheses of Section 2.1 exactly.
• Good views require knowing which texts describe the same scene; on Recap-DataComp two of each image's three texts are generated by an image captioner and carry information distilled from image-text training.
• The scaling law fails further out for good views, has no term for repeated data, and its exponents are poorly identified; the paper calls it an empirical summary of one recipe.

• A whitened JEPA with second-order anti-collapse learns the spectral embedding of its positive-pair kernel; shared positives make separately trained modalities coincide up to one rotation. §2.1 The Platonic Representation Theorem
• When positives differ, best-rotation misalignment is bounded by kernel mismatch divided by the spectral gap. §2.2 When Positives Differ
• Good-view JEPAs learn functions of the world state alone under a spectral gap, while bad-view JEPAs can spend r directions on nuisance when beta_1 > alpha_k, giving alignment error at least 2r. §2.3 Good-View and Bad-View JEPAs
• Next-token prediction has JEPA form with prefix-continuation positives and log-partition anti-collapse, so the paper treats it as a bad-view JEPA. §2.4 Next-Token Prediction is a Bad-View JEPA
• Trimming is the most important good-view knob: whole captions as views lose a quarter of alignment, .185 against .246, while masking, number of views, batch size, width and schedule move the mean by at most .016. §3.2 Trimming Makes Good Views, the Rest Barely Matters
• On MSR-VTT, good views raise mutual 10-NN on each side: .067 with good views on both sides, .059 with good text views only, .043 with good image views only and .039 with neither. §3.3 Good Views Help on Both Sides
• On the same corpus, good views reach next-token alignment with 42 times fewer samples. §4.2 Good Views Need 42 Times Fewer Samples
• A captioner whose only part fitted on image-caption pairs is one linear map with orthonormal rows reaches CIDEr .838 on COCO. §5.1 Captioning Without Training a Network on Image–Caption Pairs