MIT CSAIL researchers say GPT-6 Astra edited 0.016% of Qwen3.5-0.8B’s weights without gradients or training, and the model’s MATH-500 score rose from 7.9% to 48.1% in their experiment. di.gg/s/mtjrze6t