MIT CSAIL shows GPT-6 Astra weight edit boosts Qwen3.5-0.8B MATH-500 to 48.1%

Researchers at MIT CSAIL published a study on weight-space programming, demonstrating that GPT-6 Astra could modify the weights of a small open model, Qwen3.5-0.8B, to significantly improve its MATH-500 score from 7.9% to 48.1% without gradient descent or training. The modification acted as a rank-one change that removed a repetition loop hindering the model from completing answers. While independent re-runs confirmed the numeric results, analysis showed the edit successfully rescued cut-off generations rather than adding new mathematical capabilities. Open article →