AnalysisAI ModelsOctober 30, 2026

LeCun and Balestriero prove Platonic Representation Hypothesis for JEPAs

Read original source →arxiv.org

A 109M JEPA-Text model matched next-token prediction on WIT-1024 with 42x fewer samples (.109 at 21.6M vs .111 at 900M) and hit .184 mutual 10-NN with DINOv2, beating Qwen2.5-72B (.154) and Llama-3-70B (.151). The arXiv paper argues JEPAs trained on "good views" learn the same representation up to one rotation.

People · Yann LeCun

How this story unfolded

5 days · 1 report · 7 community posts · 8 of 10 shown

  1. Oct 30
  2. Nov 1
  3. Nov 4
  4. Nov 5

More stories today

Open the live feed