
Open 27B Model Rewrites Its Own Weights Without Training, Paper Says
A new arXiv preprint says an open 27B language model can diagnose its own failures and rewrite its weights through reasoning alone, improving on unseen problems without gradients or training.
A preprint posted to arXiv on November 16, 2026 describes self-rewriting, a method in which a language model uses its own reasoning to change its weights without gradients or training. The authors report that an open 27B model improved by 6.6 points on unseen problems after rewriting its own weights, and that it could repair a damaged copy of itself. They say the ability emerged only in the largest model in one family.
Self-rewriting without training
The paper says a language model is usually trained once and then frozen. To improve it, developers typically run more training, add prompts or harnesses, or spend extra computation at inference that ends with the task. The authors propose self-rewriting as a third option: the model acts as its own optimizer and rewrites its weights through reasoning alone. They contrast this with earlier work in which a frontier model edited a much smaller model. In this study, an open 27B model was given its weights, graded answers and the ability to run itself, then diagnosed its own failure and rewrote the weights behind it.
Testing edits and repairing damage
The authors report that the model surpassed itself: its own edit made it 6.6 points more accurate on problems it had never seen. They also say it keeps getting better by testing edits on practice problems and submitting only those that work, with each edit improving on the last; over eight consecutive sessions, its accuracy only rose. In another test, the model repaired a damaged copy of itself. The copy solved almost nothing, yet with no retraining and without reading the original weights, the model brought it back to over 80 percent of the original accuracy on image tasks it had never seen.
Scale and future implications
The ability appears to emerge with scale. In one model family, the smaller models made no edit or broke themselves, while only the largest improved and repaired itself. The authors say self-rewriting writes its gain into the weights at no cost per answer, unlike test-time scaling, and runs only forward passes, like serving, unlike training. They argue that the wall between training and inference is a limit of current tools rather than a property of intelligence, and that self-rewriting is a necessary step toward recursive self-improvement. They suggest a model could keep upgrading itself for life without a larger model, harness or prompt engineering, or pre-training or post-training.
Sources: arxiv.org
SiTech — AI-powered web development
We build fast, modern websites and bring AI into real business workflows. Have a project or a question? We'd love to help.