[MIT Research Team: GPT-6 Astra Improves Mathematical Performance by Modifying Model Weights] According to a paper published on November 3rd by the MIT CSAIL research team, researchers have improved the mathematical performance of the GPT-6 Astra model by directly modifying the weights of Qwen3.5-0.8B without using gradients or training. The research team stated that Astra modified 0.016% of the model's weights, resulting in an increase in the MATH-500 score from 7.9% to 48.1% under the same sampling settings, and an approximately 3-fold improvement in instruction compliance test IFEval performance. The team calls this method of directly reading and modifying neural network weights "weight space programming."