Compliance scores by model rose 41.2%p and 34.5%p… first-token limiting and examples used together

Harvard, MIT, and Cambridge Boston Alignment Initiative researchers reported in an October 29 arXiv paper, "The First Few Tokens Matter: Understanding Chain-of-Thought Controllability," that constraining the first token of a reasoning model's chain-of-thought while also providing user examples raised format-instruction compliance scores by 41.2 percentage points and 34.5 percentage points in two inference models, tested on Qwen3-14B and 32B among others. The compliance score measures how well the model's generated reasoning text follows formatting instructions, a metric distinct from final-answer accuracy. The finding matters because it shows that the opening tokens of a chain-of-thought disproportionately govern how controllable a model's reasoning is, offering a cheap lever for steering model behavior without retraining. What to watch next is whether the technique generalizes across other model families and whether the compliance gains translate into more reliable downstream outputs in production settings.