Harvard and MIT paper finds reasoning models' chains of thought are hard to control partly because they start with "Okay"
A new paper from Harvard and MIT finds that reasoning models control their chains of thought poorly, partly because they habitually start with "Okay." The authors had models such as Qwen3-32B reason in uppercase or number each sentence; Qwen3-32B and OLMo-3.1-32B put about 99% of the first-token probability on "Okay"; after banning "Okay" and "Alright" and adding 1 example, the average compliance rate rose by 34 to 41 percentage points compared with instructions alone. Tasks involving hidden information such as avoiding a certain word showed more unstable responses, and across 12 models this appears to be a separate capability.
Why read this
AI score 55
The original article breaks down the reasons why reasoning models struggle to control their chains of thought into the specific element of opening words, and provides reproducible intervention effects.
Topics
Body
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.