Same question, run two ways. A single forward pass guesses the answer in one leap. Give the model room to write the steps and it composes the answer instead.
pretraining
predict the next word
post-training
follow + prefer
answer-time
spend compute thinking
prompt
23 apples. They used 20 to make lunch and bought 6 more. How many apples do they have?
output
The model never changed — only whether it was allowed to write its steps. One leap is a guess; the steps let it carry 23 - 20 = 3 forward into 3 + 6 = 9. More compute at answer-time → better answers on hard problems.