Researchers report that an open 27-billion-parameter language model improved its accuracy on unseen problems by rewriting its own weights, without gradients or additional training. The reported gain was 6.6 points—and the authors say the improvement was written into the model rather than confined to a particular prompt. Source
The findings appear in Language Models Can Rewrite Their Own Weights, a preprint by Jiajun Fan and 10 coauthors submitted to arXiv on November 16, 2026, and revised on November 18. They call the approach “self-rewriting.” The researchers gave the model its own weights, graded answers and the ability to run itself. This wasn't an ordinary chatbot spontaneously discovering the controls. Source
What the model reportedly changed
Weights are the numerical parameters that shape a model's behavior. The authors say the model used its own reasoning to diagnose a failure and rewrite the weights behind it. Their procedure used only forward passes, the computations used to run a model. It didn't use gradients: mathematical signals that guide weight adjustments during training. That doesn't mean the underlying model was never trained, or that the editing experiment required no computation. Source
The model also tested edits on practice problems and submitted only those that worked. Across eight consecutive sessions, its accuracy only rose, the authors report. Source
Within the single model family described, smaller versions either made no edit or broke themselves; only the largest improved and repaired itself, according to the abstract. Editing privileges aren't a qualification. Source
The authors also say the model repaired a damaged copy of itself without retraining or reading the original weights, restoring it to more than 80% of its original accuracy on unseen image tasks. That's recovery relative to its original performance, not an absolute accuracy score above 80%. Source
The infinite-upgrade button remains unproven
The authors believe self-rewriting is a necessary step toward recursive self-improvement: a model repeatedly upgrading its own capabilities. They suggest a model could keep improving throughout its lifetime. The reported eight-session run doesn't establish that. Source
The abstract doesn't identify the benchmark behind the 6.6-point gain or provide statistical uncertainty for it. Without those details, it's hard to judge how broadly the improvement holds. Source
Comments