Can AI repair another AI's brain? MIT researchers use GPT-6 Astra to improve the math ability of Qwen3.5-0.8B

MIT CSAIL researchers had OpenAI's GPT-6 Astra directly read and write the weights of Alibaba's Qwen3.5-0.8B through a Python environment, fixing, without retraining or gradients, the problem of the model falling into repetitive reasoning, sometimes up to the 32768 token limit. According to the paper, Astra changed only 131072 weights, 0.016% of the total 800 million weights; Qwen's score on MATH-500 rose from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%, and IFEval from 21% to 62.3%. The researchers say Astra did not teach Qwen new mathematical concepts, but removed a fault that prevented it from using its existing abilities. This technique, called Weight-Space Programming, still has limitations: Astra cannot tell from weights alone which of the two models was trained to cater to users, and needs attention information. The team has released the experimental logs and is developing related benchmarks.