MIT CSAIL has introduced weight-space programming, allowing GPT-6 Astra to directly rewrite the weights of Qwen3.5-0.8B

A team of MIT CSAIL PhD students led by Rikiya Takehi has introduced weight-space programming. Using GPT-6 Astra, they can read and write neural network weights without training or gradients. By modifying only about 0.016% of the weights in Qwen3.5-0.8B, performance on MATH-500 jumped from 7.9 to 48.1, GSM8K rose from 13.8 to 67.1, and instruction following improved threefold.