true-jepa hits 89.7% unsupervised for new SOTA, author targets 100%

MarcJSchmidt · x · 2026-11-22

Marc Schmidt reports progress on true-jepa, a JEPA-style self-supervised variant: on a new SSLvMB benchmark run the new model scores 83.8%, while true-jepa xD reaches 89.7%, a new SOTA, and another version is at 94.5% — all fully unsupervised and still undertrained.

Responding to a question, he says he's confident of reaching 100%, and that the approach may be general enough to work on pixels as a world model or policy for tasks like Push-T — though that remains unverified.

Original post →

More from Research

Research channel →