MIT: GPT-6 Astra can directly modify Qwen weights, math capability surges 6-fold

MIT experiments show that GPT-6 Astra, without gradient training, changed only 0.016% of the weights in Qwen3.5-0.8B and raised its MATH-500 score from 7.9% to 48.1%, while instruction-following ability also improved 3-fold. The team calls this method Weight-Space Programming: a large model analyzes neural network weights, locates problems, and modifies them as required—for example, reducing the proportion of MATH-500 cases hitting the 32768-token length limit from 93.0% to 33.1%. Astra also identified, by analyzing weights, the games and languages a small network had learned, found induction heads, and in a small-car hill-climbing task changed 3.5% of the weights so that all 1024 tests succeeded. In MinAtar experiments, networks whose weights it wrote from scratch scored 517.4 on average, about 20 times the 25.3 of a PPO baseline trained for 500 million steps. The article also notes that Astra's capabilities still come from its prior training, and that analyzing weights and evaluation also consume computational resources.