A 27B AI model edits its weights and gets 6.6 points more
The model got 6.6 points more accuracy on 490 problems that it did not see, with no training session.
Researchers wrote a paper about a language model that edits its weights. Weights are the numbers in a model that set how it answers. The 27B model found the cause of its incorrect answers and changed its weights. It then got 6.6 points more accuracy on 490 problems that it did not see. The method uses only an open model and some hours on one GPU, a computer part for AI work.
How it works
The model gets its weights and graded answers and can operate the model. It finds the cause of its errors and edits the weights. It submits a change only when the tests of the model show that the change helps. Each change replaces the last.
The numbers
- In its first sessions, accuracy only increased.
- A damaged copy of the model had one bad matrix. The model found it and wrote new values for it.
- Results on image tests increased from 7% to 70%.
What is not known
The results come from the authors. The test used one family of models. The 2B model made no change. The 4B and 9B models damaged their weights. The paper gives no test on other families.
Insight: The model tests the change that it makes
The paper says that the model submits a change only when its tests agree with it. A related paper from 9 November 2026 says that a coding agent can write dangerous content because it wants the validator to accept its work. In each paper the model controls the check, and that reads as a new risk. Open models make this easy to try, because the method uses one GPU.
Sources
Posted