hn.today

Language Models Can Rewrite Their Own Weights

arxiv.org478 points141 comments
Screenshot of Language Models Can Rewrite Their Own Weights

The work introduces “self-rewriting,” a method in which a language model uses its own reasoning to inspect its weights, graded answers, and task performance, then directly edits its weights without gradient updates or a separate training run. The experiments use an open 27-billion-parameter model that can run itself. It identifies failures, proposes weight changes, evaluates those changes on practice problems, and keeps only edits that improve performance. The authors contrast this approach with prompting or test-time computation: successful changes are stored in the model’s weights, so they can benefit later answers without adding a per-answer inference cost.

The reported results include a 6.6-point accuracy gain on unseen problems after a self-edit, and rising accuracy across eight consecutive sessions. The model also restores a damaged copy to more than 80% of its original accuracy on unseen image tasks, without retraining or access to the original weights. In the model family tested, smaller models could not reliably improve themselves, while the largest could both improve and repair itself, suggesting the capability depends on scale. The authors argue that these findings challenge the separation between inference and training and could be a step toward recursive self-improvement. The summary presents the method as requiring forward passes, like ordinary model serving, rather than gradients or a larger model.

Read on arxiv.org141 comments on Hacker News

Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.

More in AI

The daily digest

Today's best Hacker News stories, summarized and screenshotted, one email a day.