Featured Post

An Open 27B AI Model Rewrote Its Own Weights and Gained 6.6 Points on Unseen Problems

Glowing cyan artificial brain made of network nodes on a white background

A new arXiv paper reports that an open 27B language model rewrote its own weights and became better at problems it had never seen. With no gradients and no training, the model's own edit made it 6.6 points more accurate on 490 unseen problems, rising from 57.2 percent to 63.8 percent. The authors call this self rewriting, and they say the model is its own optimizer.

The result is early research, not a product, and it only worked at the largest size the team tested. Smaller models in the same family either made no edit at all or broke themselves, so the ability seems to emerge only with scale. But the paper shows a third path beyond training and prompting.

What happened

The paper, titled Language Models Can Rewrite Their Own Weights, was submitted to arXiv on 16 November 2026 by Jiajun Fan and ten coauthors, and revised on 18 November. The team gave the 27B model its own weights and graded answers, then asked it to diagnose its own failure and write new values into the weights behind it. The model did this by reasoning alone, using only forward passes.

The gain held up on a test fixed in advance. On 490 held out problems with eight answers each, the best edit raised accuracy by 6.6 points, and a second session's edit raised it by 4.0 points, both with a p value below 0.00001. Over eight consecutive sessions, the model tested each edit on practice problems and submitted only what worked.

The single biggest edit was surprisingly small. It raised the time step bias of every linear attention layer by 0.5, which stopped the model from thinking until it ran out of tokens at the 16,384 token limit. The model had often known the right answer but kept reasoning past the limit.

The team also tested self repair. They damaged a copy by replacing one matrix with extreme random values, and the model found the corrupted matrix and rewrote it without ever reading the undamaged weights. The damaged copy went from 7 percent to 70 percent on six image benchmarks it never saw, reaching 83 percent of the undamaged model's score.

Model sizeWhat happened when it tried to edit itself
2BMade no edit
4BBroke itself
9BBroke itself
27BImproved by 6.6 points and repaired a damaged copy

Why it matters

Training and prompting both have a cost that never goes away. Training needs another full run, and a prompt vanishes when the task ends, but a self written edit stays in the weights at no cost per answer. That is why the authors say the wall between training and inference is a limit of our tools, not a property of intelligence. See also our post on the Claude Riemann Hypothesis proof claim.

The repair result is the most striking part. A model that can find and fix a corrupted part of itself points toward systems that maintain themselves instead of waiting for engineers to retrain them. The authors frame it as a step toward recursive self improvement.

Keep the limits in view. The paper tested one model family, the gains came from fixing a looping problem the model already had, and the edit did not teach it new knowledge. Eight sessions is a long way from a model that redesigns itself without bound.

What to watch

Watch whether other labs reproduce it. The paper is a preprint, and the test will be whether independent teams get the same gain on other models and tasks. If the effect is real and general, expect fast follow up work on self editing as a standard tool.

Watch the safety conversation too. A model that can rewrite its own weights is powerful, and the same ability could in principle remove a safeguard. For now, the model only improved on a fixed test, under close watch.

Read the paper

No official video covers this paper yet. The full abstract and the paper itself are on arXiv at the Language Models Can Rewrite Their Own Weights page.

Sources

arXiv: Language Models Can Rewrite Their Own Weights, arXiv 2611.21885, revised 18 November 2026

AI Info: Language Models Can Rewrite Their Own Weights, paper summary

Follow Learn Daily AI on Instagram at @learningdailyai for one new AI story every day.

Comments