AI & Tech··1 sources verified
Preprint Shows a 27B Language Model Can Edit Its Own Weights to Improve Without Training
Summarised by Relevant News AI · Read time: 3 min
A new arXiv preprint (2611.21885) reports that a 27B open model can diagnose its own failures and rewrite its weights using only its own reasoning, with no gradients or training runs. The self-edit raised accuracy by 6.6 points on 490 unseen problems, and the model repaired a corrupted copy of itself, lifting image-benchmark accuracy from 7% to 70%. The effect appeared only at the 27B size; the 2B model made no edit and the 4B and 9B models broke themselves.
Why it matters: If self-editing holds up beyond this setup, improving a deployed model could shift from costly retraining toward lightweight, persistent edits, though the results are unreviewed and limited to one model family.
All sources