AI & Tech··1 sources verified
MIT Researchers Show an LLM Can Edit Another Model's Weights to Fix a Repetition Bug
Summarised by Relevant News AI · Read time: 3 min
MIT researchers report that OpenAI's GPT-6 Astra edited about 0.016% of a small Qwen model's weights, with no gradient descent or training data, removing a habit of repeating until it ran out of context. The model's MATH score rose from 7.9% to 48.1%, but most of the gain came from fixing the repetition: math answers hitting the length limit fell from 93% to 33.1%, not from new math knowledge. The three arXiv papers from November 6 are still awaiting independent reproduction, and Matthew Honnibal has argued the claims are overstated.
Why it matters: If weight-level repairs hold up, model debugging could start to resemble patching code, which would change how teams maintain and fix deployed models.
All sources