Harvard 与 MIT 论文：推理模型控制不了思维链，问题出在开头总说 Okay
模型思维链开头总是 Okay，导致它不听你的推理指令。这篇论文发现禁掉两个开头词就能提升 34 个百分点，很实用的细节。
Harvard 与 MIT 的论文研究了推理模型为何难以按指令控制自己的思维链。作者让 Qwen3-32B、OLMo-3.1-32B 等模型用大写或逐句编号来推理，发现这些模型把约 99% 的首词概率放在 Okay 上。实验中禁止 Okay、Alright 两个开头词并加入 1 个示范样例，使模型在 12 个模型上的平均遵循度比单纯指令提升 34 到 41 个百分点。像避免说出某个词这类需要隐藏信息的任务提升有限，属于另一类能力。论文建议把开头词问题与整体可控性分开报告，避免安全团队高估思维链监控的可靠性。
New Harvard and MIT paper finds that reasoning models' poor control over their chain-of-thought partly comes from a habit of opening it with "Okay".
Safety teams read a model's reasoning to catch misbehavior, and benchmarks showing models can't steer that reasoning have been reassuring. Why models fail at this was unclear.
The authors told models like Qwen3-32B to reason in uppercase or number each sentence. Qwen3-32B and OLMo-3.1-32B put about 99% of first-token probability on "Okay", so the authors tried banning it or prefilling 1 compliant token.
Banning "Okay" and "Alright" and adding 1 worked example raised average compliance by 34 to 41 percentage points over instructions alone. Tasks that hide information, like avoiding a word, responded less consistently and look like a separate skill across 12 models.
If you trust chain-of-thought monitoring because of controllability scores, separate the opening-word problem from general control and report hiding tasks separately.
