Artificial Intelligence1 min read

Language Models Can Rewrite Their Own Weights

By · Published by Everything Blog

In short

Artificial intelligence researchers have demonstrated that a language model can improve its performance by rewriting its own weights through reasoning, operating without traditional training, gradients, or external assistance. In the study, an open 27-billion-parameter model was provided with its own weights, graded answers, and the ability to execute its code. By testing modifications on practice tasks and retaining only successful adjustments, the model steadily increased its accuracy across eight consecutive se…

Key points

  • Recursive self-improvement can be achieved through pure forward passes without gradients…: Recursive self-improvement can be achieved through pure forward passes without gradients or reliance on larger external models.
  • The capability to reliably rewrite internal weights emerges with scale, whereas smaller m…: The capability to reliably rewrite internal weights emerges with scale, whereas smaller models in the same family damaged their parameters.
  • Direct parameter modification provides a viable path to permanently embed performance gai…: Direct parameter modification provides a viable path to permanently embed performance gains rather than relying on prompt engineering or retraining cycles.

Sources checked by our analysis pipeline: 1

Computer Science > Artificial Intelligence

[Submitted on 16 Nov 2026 (v1), last revised 18 Nov 2026 (this version, v2)]

Title:Language Models Can Rewrite Their Own Weights

View PDF HTML (experimental)Abstract:A language model is trained once and frozen. Making it better takes another training run, or a prompt, a harness or extra computation at inference that is gone once the task ends. We show a third way: to our 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. We call this self-rewriting: the model is its own optimizer. In earlier work, a frontier model edited a much smaller one; here an open 27B model, given its weights, its graded answers and the ability to run itself, diagnoses its own failure and rewrites the weights behind it. It surpasses itself: its own edit makes it 6.6 points more accurate on problems it never saw. It keeps getting better: it tests its edits on practice problems and submits only what works, each edit improving on the last,…

Artificial Intelligence