Language Models Can Rewrite Their Own Weights Without Training

A new arXiv paper shows a 27B language model can rewrite its own weights without training or gradients, boosting unseen-task accuracy by 6.6 points and repairing a corrupted copy from 7% to 70% accuracy.

2026-11-18 ~ 2026-11-18 · 3 related posts

1 near-duplicate retellings: jiqizhixin