GPT-6 Astra reportedly boosts Qwen3.5-0.8B’s MATH-500 score by editing 0.016% of its weights

MIT CSAIL researchers report that GPT-6 Astra improved Qwen3.5-0.8B’s MATH-500 score by editing just 0.016% of its weights. They call the approach Weight-Space Programming and say Astra worked without gradients or training loops, though it used a Python kernel to make the edit. They also report a threefold IFEval gain. The performance gains and the claimed diagnosis of reasoning loops are the researchers’ experimental findings.