What Qwen is thinking when it gets stuck saying "Wait"
MIT CSAIL had GPT-6 Astra fix Qwen3.5-0.8B's reasoning loops by editing 0.016% of its weights (MATH-500 went from 7.9% to 48.1%). They released the diff, so the edit can be rebuilt exactly and run locally. I took it a step further and had Claude look at Qwen with Anthropic's Jacobian lens.