The MIT CSAIL team had GPT-6 Astra directly modify the weights of Qwen3.5-0.8B, raising the MATH-500 score from 7.9% to 48.1%

The MIT CSAIL team had GPT-6 Astra directly read and modify the weights of Qwen3.5-0.8B through natural language, and by changing only about 0.016% of the parameters, it raised the MATH-500 score from 7.9% to 48.1%. Within a limited number of forward computations, Astra located a loop direction written into the MLPs of layers 15 and 19, subtracted that direction from each of the 64 output columns of these two layers, and modified a total of 131,072 weights, without retraining or using gradient descent. On Qwen3.5-0.8B, GSM8K rose from 13.8% to 67.1%, IFEval rose from 21.0% to 62.3%, and the proportion hitting the 32,768 Token limit fell from 93% to 33%. Blog address: https://rikiyat.github.io/blog/weight-space-programming/