MIT's weight-space programming lets GPT-6 Astra rewrite Qwen3.5-0.8B weights directly

On November 2, 2026, MIT CSAIL researcher Rikiya Takehi published a research effort on his personal blog called Weight-Space Programming. The idea is to let a large language model read, understand, and edit another neural network's weights directly, rather than adjusting parameters through a standard training pipeline. The same discussion appeared on Hacker News, where the summary said Astra can directly read and write model weights to improve Qwen in weight space instead of relying only on fine-tuning or prompting; the post drew 630 points and 182 comments, with the conversation focused on whether the method can be reproduced and whether the weight edits are stable.

On November 3, QbitAI reported that an MIT study showed GPT-6 Astra could directly modify Qwen's model weights, with mathematical ability up 6x and instruction-following ability also improving, though the material did not provide paper details or experimental setup. The same day, IT Home added concrete experiment data. The research, co-authored by Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab, used OpenAI's GPT-6 Astra to analyze and modify the neural network weights of Alibaba Cloud's Tongyi Qianwen model Qwen3.5-0.8B. By adjusting only about 0.016% of the parameters, it raised the model's score on the MATH-500 benchmark from 7.9% to 48.1%, about 6.1x the original, and the process reportedly used no gradient descent or traditional training loop. IT Home also said GPT-6 Astra did not simply tweak a small number of parameters at random; it first located Qwen3.5-0.8B's reasoning-loop problem, with the model prone to repeatedly generating the same kind of steps when answering math questions.

The reports differ on which model was involved. The Hacker News summary only mentioned improving Qwen, while QbitAI referred to modifying Qwen's weights, and IT Home specified Qwen3.5-0.8B. The public record currently stops at the November 3 blog post and media coverage, and the material does not provide paper details or experimental setup.