Language Models Can Rewrite Their Own Weights
Summary
The arXiv paper 'Language Models Can Rewrite Their Own Weights' introduces self-rewriting, where a 27B language model uses its own reasoning to edit its weights without gradients or retraining, improving accuracy on tasks it hadn’t previously seen. The model can diagnose its failures, perform edits that survive across sessions, and even repair damaged copies, with performance gains emerging as models scale.