Platonic Representation Hypothesis Proven: 109M JEPA Aligns with Vision Better than 72B LLMs
ClementineDomi6 · x · 2026-10-30
- Randall Balestriero (with Yann LeCun) announces the Platonic Representation Hypothesis is now a theorem: JEPAs trained separately on any modalities provably learn the same representation up to one rotation.
- Proof intuition: 'good views' preserve shared world structure while independently resampling modality-specific nuisance; with whitened MSE and a spectral gap, models converge to the same world eigenspace in different orthonormal bases, so a rotation suffices to align them.
- The companion paper λ-JEPA (arXiv:2609.35288) introduces SACReg, a spectral anti-collapse regularizer based on λ-balance, addressing dimensional collapse in the backbone; it beats LeJEPA and VISReg on ImageNet-1k classification and linear-probe transfer across eight datasets, and improves on V-JEPA 2 for video SSL.
- Headline result: a 109M JEPA-Text model aligns with vision models better than 72B LLMs.
More from Models
- Arena weekly: Gemini 4 Carbon tops Text Arena, GPT-6.1 Terra enters Agent Arena at #8 — arena · 2026-10-31
- 6 October models, same 5 trap-laden agent tasks: DeepSeek does it all for $0.02, hyped Flint passes 1/5 — latent_lurker8587 · 2026-10-31
- Security researcher: air-gapped AI agents are only possible with open-weight models — wunderwuzzi23 · 2026-10-31
- Gemini 4 Carbon tops Artificial Analysis index at 59, 23% cheaper per task than Opus 5.5 — rohanpaul_ai · 2026-10-31
- Podcast: Is Mistral Large 4 the West's best open model? Plus AI cyber threats — The Cognitive Revolution · 2026-10-31
- OpenAI shows GPT-6 Astra remotely controlling a robot arm in builder community — OpenAIDevs · 2026-10-31