arXiv:2611.21885v1 Announce Type: new Abstract: Language models are trained once, frozen and deployed: improving one again takes another training run, and a prompt, a scaffold or extra computation at inference is gone once the task ends. A frontier model has been shown to edit the weights of a much smaller one. We ask whether a model can rewrite its own weights to make itself better, and find that it can, by its own reasoning, with no gradients and no training; we call this self-rewriting. Given its weights, graded answers and the ability to run itself, a 27B model diagnoses its failure and edits the weights behind it. It surpasses itself: its own edit makes it 6.6 points more accurate on 490 problems it never saw, in a test fixed in advance, and another session's edit also makes it significantly better. It keeps getting better: it submits an edit only when its own tests support it, each edit replacing the last, and over its first eight such sessions its accuracy only rises. Unlike test-time scaling, the gain persists in its weights and adds no cost per answer; unlike training, it needs no pre- or post-training infrastructure, only an open model and hours on one GPU. It repairs a damaged copy of itself: it locates the corrupted matrix and rewrites it with values of its own, never reading the undamaged weights; found on text math problems alone, the repair lifts the copy on six image benchmarks it never saw from 7% to 70% accuracy, 83% of the undamaged model's. And the ability emerges with scale: across four sizes of one family, the 2B model makes no edit and the 4B and 9B models break themselves; only the 27B model works on itself as a human would, comparing its hidden states on answers that finish and on answers that run out of tokens, weakening the direction that separates them. Like a human, a model need not separate training from inference; improving a language model may come to take a conversation rather than a training run.
Language Models Can Rewrite Their Own Weights
A 27B-parameter language model can diagnose its own failures and rewrite its own weights without gradients or training, gaining 6.6 points of accuracy on 490 held-out problems, according to arXiv paper 2611.21885v1. The paper reports that the model submits an edit only when its own tests support it, and its accuracy rose across its first eight such self-rewriting sessions. The same method repaired a deliberately corrupted copy of the model, lifting it from 7% to 70% accuracy on six unseen image benchmarks, 83% of the undamaged model's score; the ability emerged only at 27B, with 2B, 4B and 9B models failing to edit themselves usefully.
A 27B model that compares its hidden states on answers that finish versus answers that run out of tokens can edit its own weights without gradients, gaining 6.6 points on 490 held-out problems and repairing a corrupted copy from 7% to 70% on six unseen image benchmarks.
Run your AI side-project on zahid.host
EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.