2026-11-03 · America/Los_Angeles · 社区动态 · #11

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. At the token where the model decides to write "Wait" again, the lens reads "Wrong", "Nope" and "Hmm" from about 60% depth on, even though the answer is already written out. At real stopping points it reads "</think>", "Answer" and "Output". With the edit applied, the lens at the same token shows…

热度 39.9 / 100；排名与评分保留该期记录。