A language model can make itself better by rewriting its own weights, according to a new preprint from a team of 11 researchers. The paper, posted to the arXiv preprint server, describes an open 27-billion-parameter model that reasons about its own mistakes and then edits the internal parameters responsible for them. The process requires no gradients and no additional training.

Today, a language model is typically trained once and then frozen. LLM optimization usually means another training run, or relying on a prompt, a surrounding software harness, or extra computation at inference time. Those gains disappear once the task ends. The authors describe a third option, which they call self-rewriting, in which the model acts as its own optimizer.

How does self-rewriting let a language model improve itself?

The researchers say an earlier study used a frontier model to edit a much smaller one. In the new work, the model is given its own weights, its graded answers, and the ability to run itself. It identifies where it failed and rewrites the weights behind those failures. The authors report that the resulting edit made the model 6.6 points more accurate on problems it had never seen.

The method also appears to build on itself. According to the paper, the model tests its edits on practice problems and submits only the changes that work. Each edit improves on the last. Over eight consecutive sessions, the model's accuracy only rose.

The team also tested whether the model could repair damage. They created a broken copy of the model that solved almost nothing. Without retraining and without ever reading the original weights, the model restored the copy to more than 80 percent of the original's accuracy on image tasks it had not seen before.

Does self-rewriting work on smaller models?

The ability seems to depend on scale. In one model family, the smaller models either made no useful edit or damaged themselves. Only the largest model improved and repaired itself, the authors write.

How does self-rewriting compare with training and test-time scaling?

The authors contrast self-rewriting with two common approaches. Test-time scaling adds computation for each answer, so it carries a cost every time the model responds. Training runs a separate process to update the model. Self-rewriting, by contrast, writes its gains directly into the weights, so it adds no cost per answer. It also runs only forward passes, the same kind of computation used when a model serves answers, rather than the backward passes used in training.

The researchers argue that the line between training and inference reflects the limits of current tools, not a fundamental feature of intelligence. They say self-rewriting is a necessary step toward recursive self-improvement, in which a model keeps upgrading itself over its lifetime. Such a system would not need a larger model, prompt engineering, or separate pre- or post-training stages, they write.

The paper is a preprint, and its findings have not been independently verified. The first version was submitted on Nov. 16, 2026, and a revised version followed on Nov. 18, 2026.

Dieser Artikel wurde mit Unterstützung von KI verfasst.
News Factory APP - agentische News für besseres SEO & AEO.