Introduction
Deployed language models are usually frozen. Improving one often means another training run, while extra prompts or inference-time computation do not permanently change the model. A paper introduces a different possibility: let a model inspect its own results, identify weaknesses, and edit its weights. The authors call this process “self-rewriting.”
Key findings
- The model diagnoses and edits itself. In the reported setup, a 27B model has access to its weights, graded answers, and the ability to run itself. It uses its own reasoning to identify failures and propose weight changes. The process uses neither gradients nor conventional training.
- One edit improves results on unseen problems. The paper reports a 6.6-percentage-point accuracy increase on a test of 490 problems that the model had not seen, with that evaluation fixed in advance. An edit produced in a separate session also reportedly improved performance significantly.
- Edits are screened before adoption. The model submits a change only when its own tests support it, and each accepted edit replaces the previous one. Across the first eight sessions, the reported accuracy rose at every step.
- The approach is also used to repair a damaged copy. The model locates a corrupted matrix and rewrites it with values of its own, without reading the intact model’s corresponding weights. Although the repair was identified using text math problems, the paper reports that accuracy on six unseen image benchmarks rose from 7% to 70%—83% of the undamaged model’s performance.
Why it matters—and what remains unknown
If replicated across tasks and independent evaluations, self-rewriting could offer a route between prompt-based adaptation and full retraining. Unlike an inference-time trick, an accepted change persists in the weights and, according to the authors, adds no per-answer inference cost. They also describe the setup as requiring an open model and hours on one GPU, rather than pre- or post-training infrastructure.
The findings do not yet establish that models can safely or generally improve themselves. The available summary describes particular experiments but does not give enough detail to judge how well the method transfers across architectures, tasks, or model sizes. Nor does it show that every proposed edit is reliable. Because the model helps generate and screen its own changes, independent test sets, long-term regression checks, and ways to roll back harmful edits will matter. The key question is not simply whether a model can alter its parameters, but whether it can do so consistently under verifiable safeguards.
Source: Hugging Face Daily Papers
Comments
Checking sign-in status...
Loading comments...