The paper teaches AI models to plan their thinking time like a smart test-taker who has to finish several questions before the bell rings.
The paper shows that making a model write a number as a sequence of digits and then grading the whole number at the end works better than grading each digit separately.