The Qwen weight edits are fixing the repetition and works on non math as well

An MIT CSAIL team had an AI agent edit 0.016% of the weights of Qwen3.5-0.8B, raising its MATH-500 score from 7.9% to 48.1% with thinking on. A follow-up analysis finds the edit mainly removes a direction tied to repetitive thinking loops, not math skill: it also stops loops on non-math…