StudioMIT researchers use GPT-6 Astra to directly rewrite Qwen3.5-0.8B weights, lifting MATH-500 from 7.9% to 48.1%
Original title: MASSIVE. MIT researchers showed that GPT-6 Astra can improve another AI model by changing its weight…
MIT researchers had GPT-6 Astra improve Qwen3.5-0.8B's math ability by directly editing its weights, with no training or gradients. Editing just 0.016% of the weights raised MATH-500 accuracy from 7.9% to 48.1%. Given the model's weights in a Python notebook and a budget of 50 test runs on practice math problems, Astra found the model repeated itself until hitting its length limit, traced the habit to a single internal signal written by 2 layers, and subtracted that signal from the weights, letting the model finish answers it could already solve. The team calls the approach Weight-Space Programming.
The relayed account includes concrete experiment details such as the weight edit ratio and before/after scores, offering readers a transferable look at the weight-space programming method.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.