Papers & Research
Weight-space programming
MIT CSAIL: GPT-6 Astra lifts Qwen3.5-0.8B's MATH-500 from 7.9% to 48.1% by editing its weights directly
Researchers gave GPT-6 Astra a Python session over the weights of Qwen3.5-0.8B, 50 forward passes and no gradients or training run. Astra changed 0.016% of the weights (131,072 numbers in two MLP layers): MATH-500 went from 7.9% to 48.1% and IFEval from 21.0% to 62.3%, and Hugging Face's akhaliq independently rebuilt the checkpoint at 51.0%. The base model hit the token limit on 93% of its answers, so fixing that looping bug explains most of the gain.
7.9% → 48.1% MATH-5000.016% of the weights changed21.0% → 62.3% IFEval