Blocking Common Opening Tokens and Adding an Example Raised Reasoning Compliance in Two AI Models.

Blocking two models’ habitual opening tokens and supplying an example raised compliance scores by 41.2 percentage points for Qwen3-32B and 34.5 percentage points for OLMo-3.1-32B-Think on written-reasoning instructions. The scores measured instruction following, not final-answer accuracy.

Qwen3-32B’s average score rose from 7.8% to 49.0%, while OLMo-3.1-32B-Think’s increased from 2.9% to 37.4%. The combined intervention blocked “Okay” and “Alright” at the first reasoning position and included an example. The results covered eight control tasks.

Without instructions controlling their openings, “Okay” accounted for 99.89% of the first-token probability in Qwen3-32B and 99.52% in OLMo-3.1-32B-Think. “Okay” and “Alright” together made up more than 99.99% in both models. The paper’s authors, Hugh Van Deventer, Dmitrii Troitskii and Laura Ruis, describe this concentration at the start of some reasoning traces as an “entropy collapse.”

The tasks included adding markers or numbers to sentences, changing sentence endings, writing in Spanish or uppercase, using leetspeak, avoiding first-person language and omitting a specified word. The researchers found more varied results on suppression tasks, such as leaving out a word, than on tasks involving added information or style changes.

The reported gains came from a specific combination tested on these two models; they do not establish the same effect across reasoning models generally. The preprint also compares instruction compliance and reasoning-trace monitorability across 12 models. Those comparisons do not show that the opening-token intervention improved monitoring or answer accuracy.

Reasoning traces are text some AI models produce before giving a final answer. Researchers have examined them as one possible signal for monitoring model behavior, though they do not establish whether a final answer is correct. The findings add to research on monitoring models’ intermediate reasoning.

The preprint, “The First Few Tokens Matter: Understanding Chain-of-Thought Controllability,” was submitted Oct. 29, 2026, at 4:14 p.m. ET (20:14 UTC).