LeCun and Balestriero prove the Platonic Representation Hypothesis: JEPA views beat scaling with 42x fewer samples
burny_tech · x · 2026-11-01
Yann LeCun and Randall Balestriero (Brown/NYU/AMI Labs) turned the Platonic Representation Hypothesis into a theorem.
Key results:
- A Platonic Representation Theorem: JEPAs trained separately on any modalities learn the same representation up to one rotation when their positive pairs are two observations of one scene ("good views") — no image-text pairs needed.
- Next-token prediction is shown to be a JEPA with "bad views" (parts of one document), explaining why LLMs align with vision only through scale.
- Validated on 145k captions up to 940M web-image texts, uncovering an alignment scaling law that predicts unseen runs within seed noise.
- Sample efficiency: good views reach the same alignment with 42x fewer samples than next-token prediction; a 21.6M-sample text JEPA matches next-token training on 900M samples and outperforms Qwen2.5-72B and Llama-3-70B (trained on 15-18T tokens).
- A 109M-parameter text JEPA aligns with DINOv2 and CLIP image encoders better than every evaluated LM up to 72B, and better than CLIP's text encoder.
Showcase: a single rotation turns separately trained encoders into an image captioner (CIDEr 0.838 on COCO), with no pairs; on a pair-free benchmark the text JEPA hits 64% top-5 zero-shot on ImageNet-100, beating all reported LMs up to 8B.
Takeaway: positive views, not scale alone, are the lever for platonic representations.
More from Research
- Control Vectors experiment: letting an agent manage its own activations — sloppenheimer · 2026-11-01
- Claude designed scorpion toxin binders: 17 of 211 lab-tested designs bound — shoumikgoswami · 2026-11-01
- Hugging Face open-sources Carbon DNA foundation models with full training recipes — LoubnaBenAllal1 · 2026-11-01
- Carbon-X: midtrained Qwen3-14B beats every open LLM on DNA benchmarks — LoubnaBenAllal1 · 2026-11-01
- RL fails without midtraining: Qwen3-14B stays at chance on FrontierVEP — LoubnaBenAllal1 · 2026-11-01
- FineDNA midtraining lifts DNA scores without degrading English — LoubnaBenAllal1 · 2026-11-01