One AI rewrote another AI's brain — and its math score jumped sixfold

MIT CSAIL researchers handed OpenAI's GPT-6 Astra the weights of Alibaba's small open model Qwen3.5-0.8B and asked it to edit them directly — no training run, no gradient descent. Astra changed just 0.016% of the weights, 131,072 out of roughly 800 million.

The edited model's MATH-500 score rose from 7.9% to 48.1%, and the share of answers that ran to the token limit fell from 93% to 33.1%. Astra traced the gain to a numerical direction in layers 15 and 19 that made the small model loop instead of answering, and subtracted it.

The catch, flagged by critics and the authors' own follow-up: on Gemma 3 270M, which rarely loops, Astra's edits produced no established improvement. This looks like a powerful debugging trick — an AI patching another AI's flaw — not proof that one model taught another new math.