AI preprint links limited control over written reasoning to opening tokens

An Oct. 29 preprint links AI models' limited control over written reasoning to low variability in the first tokens after a thinking tag. Prefilling even one token substantially increased controllability; banning frequent openings alongside in-context examples raised average compliance by 34 to 41 percentage points across models. Gains were concentrated in adding or restyling information, while the ability to leave information out varied more.