GPT-6 Astra directly rewrites Qwen3.5-0.8B weights, MATH-500 score rises from 7.9% to 48.1%

The MIT CSAIL research team published the results of an experiment called "weight space programming": GPT-6 Astra directly modified 131,072 weights in Qwen3.5-0.8B via natural language instructions, accounting for 0.016% of all parameters, and the model's MATH-500 score rose from 7.9% to 48.1% as a result. The whole process took 20.3 hours, including 158 rounds of dialogue, 80 Python executions and 50 model evaluations; GSM8K rose from 13.8% to 67.1%, IFEval rose from 21.0% to 62.3%, and the proportion of responses that hit the 32,768 token limit because they fell into repetitive loops fell from 93% to 33.1%. The team said Astra did not inject new mathematical ability, but only removed obstacles preventing the model from using abilities it already had; when the same experiment was repeated on Gemma 3 270M, the MATH-500 score fluctuated by no more than 1 percentage point.