MIT Research Team: GPT-6 Astra Modifies Model Weights to Improve Math Performance

On November 3, according to an article published by the MIT CSAIL research team, researchers used GPT-6 Astra to directly modify the weights of Qwen3.5-0.8B without using gradients or training to improve its mathematical capabilities.

The research team noted that Astra modified 0.016% of the model's weights, increasing the MATH-500 score from 7.9% to 48.1% under identical sampling settings, and improving performance on the instruction-following benchmark IFEval by approximately threefold. The team refers to this method of directly reading and modifying neural network weights as "weight space programming."