ResearchNov 20, 2026, 11:29 AM EST

Researchers Report a 27B Model That Rewrites Its Own Weights

Jiajun Fan and ten co-authors posted a preprint on arXiv describing what they call self-rewriting: an open 27B language model that edits its own weights using its own reasoning, with no gradients and no training run. Given its weights, graded answers and the ability to run itself, the model diagnoses its failures and rewrites the parameters behind them. In the reported results it gained 6.6 points on problems it had not seen, kept improving across eight consecutive editing sessions, and restored a damaged copy of itself to over 80% of its original accuracy on unseen image tasks. The authors say the behavior appeared only in the largest model of the family they tested.

Why it matters

Unlike test-time scaling, the gain is written into the weights, so it adds no cost per answer; the authors argue a model could improve itself without separate training runs or prompt engineering.

Sources

Written by AIToolMaps from the source above and checked against it sentence by sentence.

Comments

No comments yet