Zoom researchers detail a “chain of draft” method to let LLMs accurately solve reasoning problems with as little as 7.6% of the tokens used by current methods.