MIT researchers use GPT-6 Astra to repair a model's weights, lifting math scores sixfold

Four MIT CSAIL researchers published a paper on November 2 describing how the GPT-6 Astra system reads and writes neural network weights directly, without gradient descent or retraining. In the key experiment, Astra modified only about 0.016% of the roughly 800 million parameters in the Qwen3.5-0.8B model, which raised its MATH-500 score from 7.9% to 48.1%. The researchers describe the work as repairing a looping failure mode rather than adding new knowledge, since the model's repetition rate dropped sharply. The search process was costly, taking 20.3 hours across 50 evaluation requests, and results were weaker on a smaller model, Gemma 3 270M. Outside commentators offered mixed views, with some noting omitted negative results and others warning about the potential acceleration of AI capabilities.