Researchers report direct weight edits reduced Qwen reasoning loops

Researchers report that GPT-6 Astra improved Qwen3.5-0.8B's benchmark results by directly editing its numerical weights without gradient-based training. With 50 evaluation runs, MATH-500 rose from 7.9% to 48.1%, while instruction-following scores on IFEval rose from 21.0% to 62.3%. The experimental method changed 131,072 weights, or 0.016%, to reduce repetitive reasoning that prevented answers from finishing. The researchers released experiment logs, code and evaluation records. They caution that removing loops exposed existing ability rather than adding mathematical skill; the method failed to establish an improvement on Gemma 3 270M, which rarely loops.