Reproduce “Language Models Can Rewrite Their Own Weights” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiajun Fan, Neil He, Changyi Yang, Andy Mao, Xi Zhu, Chaoran Cheng, Kaiqu Liang, Shuaike Shen and 3 more
Fan et al.’s method lets a language model inspect and directly edit its own weights, then test whether the change helps. In a controlled session, the model diagnoses failures and chooses edits; changes persist without computation per answer. An open Qwen3.5 27B model raised held-out math accuracy from 57.2% to 63.8%, but evidence is limited to one model family and tested sessions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.