Paper claims an open 27B model rewrote its own weights without gradients
A paper on arXiv (cs.AI, submitted 16 November 2026, revised 18 November 2026) titled "Language Models Can Rewrite Their Own Weights" makes a bold claim. Its starting point is that a language model is trained once and then frozen. Making it better normally takes another training run, or a prompt, a harness or extra computation at inference that disappears when the task ends. The authors say they show a third way: to their knowledge for the first time, a language model rewrites its own weights by its own reasoning and becomes better, with no gradients and no training. They call this self-rewriting, and describe the model as its own optimizer.
The setup, as the abstract describes it: an open 27B model is given its weights, its graded answers and the ability to run itself. It diagnoses its own failure and rewrites the weights behind it. The authors note that in earlier work a frontier model edited a much smaller one; here the model edits itself.
The abstract reports three results. First, the model surpasses itself: its own edit makes it 6.6 points more accurate on problems it never saw. Second, it keeps improving. It tests its edits on practice problems and submits only what works, each edit improves on the last, and over eight consecutive sessions its accuracy only rises. Third, it repairs a damaged copy of itself. The copy solves almost nothing, yet with no retraining and without ever reading the original weights, the model brings it back to over 80% of the original accuracy on image tasks it never saw. The authors compare this to a broken robot that fixes itself without being rebuilt.
The authors also say the ability emerges with scale. In one model family, the smaller models make no edit or break themselves, and only the largest improves and repairs itself.
They contrast the method with two existing approaches. Unlike test-time scaling, self-rewriting writes its gain into the weights, at no cost per answer. Unlike training, it runs only forward passes, as serving does. Their closing argument is that the wall between training and inference is a limit of our tools, not a property of intelligence. They believe self-rewriting is a necessary step toward recursive self-improvement: a model that, like a human, does not separate training from inference could keep upgrading itself for life, with no larger model, no harness or prompt engineering, and no pre- or post-training. That last point is a belief stated by the authors, not a demonstrated result.
Key facts
- The paper's authors claim an open 27B model, given its weights, its graded answers and the ability to run itself, rewrote its own weights with no gradients and no training.
- The reported gain is 6.6 points of accuracy on problems the model never saw; over eight consecutive sessions its accuracy only rises.
- The model also repaired a damaged copy of itself without retraining and without reading the original weights, restoring over 80% of the original accuracy on unseen image tasks.
- The authors say the ability emerges with scale: in one model family, smaller models make no edit or break themselves, and only the largest improves.
- The authors say the method writes gains into the weights at no cost per answer and needs only forward passes, and they believe it is a necessary step toward recursive self-improvement.
Why it matters
If the claim holds, it blurs the line between training and inference. Today a model is frozen after training, and gains at inference, such as extra computation or prompting, vanish when the task ends. The authors say self-rewriting keeps the gain in the weights, costs nothing extra per answer, and needs only forward passes, the same operation as serving. They frame it as a step toward recursive self-improvement, though that is their belief and not something the paper shows.
Who it affects
Researchers working on self-improving models, test-time compute and model editing are the most direct audience, since the paper positions itself against both training and test-time scaling. The abstract names no company or product, and it does not name the open 27B model or its developer, so it is not yet clear which practitioners could try this on their own models.
How to use it
There is nothing here to adopt yet. The abstract describes a research result, not a tool, and it does not describe the mechanism of the weight edit beyond saying it runs only forward passes. The practical step for now is to read the full paper, which the arXiv page offers as a PDF and an experimental HTML version, and to watch for replication.
How solid is it
This is a preprint, and the evidence here is the authors' own abstract. The abstract does not name the benchmark, the problem set or the task types behind the 6.6-point gain, nor the baseline accuracy, and it does not say what the 6.6 points are relative to or whether the gain was repeated across runs. It also does not say how many points accuracy rose over the eight sessions. The over 80% repair figure is for image tasks only, and no numeric score is given for the damaged copy. The abstract gives no peer-review status, and no independent replication or outside reaction is reported. The 6.6 points are absolute points, while the over 80% is a share of the original model's accuracy.
Risks and caveats
The "first time" claim is the authors' own, made "to our knowledge". The scale finding is limited to one model family, and the abstract does not name that family or the sizes of the smaller models that fail. Those smaller models either make no edit or break themselves, so the method does not work across the board. The recursive self-improvement framing is a belief, not a result. Until the full paper's details and independent replication are available, the numbers should be read as claims.
“The wall between training and inference is a limit of our tools, not a property of intelligence.”
— Abstract of "Language Models Can Rewrite Their Own Weights", arXiv