GPT-6 Astra Edits Qwen Weights, Raising MATH-500 to 48.1%

Published 8 November 2026

An early, non-peer-reviewed study reports that GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights, reducing thinking loops and raising MATH-500 performance from 7.9% to 48.1%. The results point to weight-space debugging rather than new mathematical ability, with failures on Gemma and sycophancy detection still limiting the claim.