Jiajun Fan

Language Models Can Rewrite Their Own Weights

In brief

What the paper does.

Self-rewriting is a third way to make a language model better: the model rewrites its own weights by its own reasoning, with no gradients and no training. To our knowledge, this is the first time a language model has done so. The model is its own optimizer. It takes no frontier model: the model’s own edit makes it 6.6 points more accurate on problems it never saw. Over eight consecutive sessions its accuracy only rises. Like a broken robot that fixes itself without being rebuilt, it repairs a damaged copy of itself without ever seeing its original weights, back to over 80% of the original accuracy on image tasks it never saw. The ability emerges with scale: of four sizes of one model family, only the largest improves and repairs itself. No training pipeline to build, no cluster to rent, no added cost per answer. The wall between training and inference is a limit of our tools, not a property of intelligence.

Published in
arXiv preprint
Paper
arxiv.org/abs/2611.21885
Project page
jiajunfan.com/projects/self-rewriting/

Cite

Cite the paper.

@article{fan2026selfrewriting,
  title         = {Language Models Can Rewrite Their Own Weights},
  author        = {Jiajun Fan and Neil He and Changyi Yang and Andy Mao and
                   Xi Zhu and Chaoran Cheng and Kaiqu Liang and Shuaike Shen and
                   Zetian Li and Ruihan Guo and Ge Liu},
  journal       = {arXiv preprint arXiv:2611.21885},
  eprint        = {2611.21885},
  archivePrefix = {arXiv},
  year          = {2026}
}