Researchers show 27B language model can rewrite its own weights, raising accuracy 6.6 points
Researchers report that a 27B model can edit its own weights without gradients or training, improving unseen math accuracy by 6.6 points and repairing a damaged copy. The result challenges the assumption that deployed models are fixed, because self-modifying weights can change behavior and auditability.
In a paper submitted Nov. 16, 2026, Jiajun Fan and coauthors report that a 27-billion-parameter language model can rewrite its own weights without gradients, fine-tuning or external training infrastructure. The model is given its weights, graded answers and the ability to run itself. It diagnoses a failure, edits the weights behind that failure, and replaces its prior edit with a newer one. The authors call the process self-rewriting and say it works by the model's own reasoning.
The paper reports a 6.6-point accuracy gain on 490 problems the model had not seen, using a test fixed in advance. Another session's edit also improved performance significantly. The model submits an edit only when its own tests support it, and over its first eight sessions accuracy rises monotonically. In a repair task, it locates a corrupted matrix in a damaged copy and rewrites it without reading the undamaged weights. The paper is posted on arXiv and not yet peer-reviewed.
The gain persists in the weights and adds no cost per answer, unlike test-time scaling, and needs no pre- or post-training pipeline beyond an open model and hours on one GPU. It extends prior work showing a frontier model can edit a smaller model. The ability appears with scale: a 2B model makes no edit, 4B and 9B models break themselves, and only the 27B model edits itself in a human-like way by comparing hidden states on finished answers and token-exhausted answers.
This matters because deployed models are usually treated as fixed artifacts that can be audited, versioned and controlled after training. If a model can modify its own weights from reasoning and test feedback, that assumption weakens: the system's behavior may no longer be fully described by a frozen checkpoint. The paper also shows transfer beyond the repair target, lifting six unseen image benchmarks from 7% to 70%, suggesting edits can affect capabilities not directly tested.
It affects developers and regulators who rely on frozen model checkpoints for auditing and safety, because self-rewriting could change a model's behavior after deployment without new training.
Sources (1)
- [1]Language Models Can Rewrite Their Own Weights(https://arxiv.org/abs/2611.21885)