Large language models can now tell when they need to reason, saving money

Large language models that use step-by-step reasoning can now figure out on their own when to use that reasoning, instead of always doing it. This makes them cheaper to run and more accurate, because they only use complex thought processes when a task actually requires them.

Large language models often waste computing power by 'thinking aloud' through every problem, even simple ones. This paper shows how models can decide for themselves when to engage in complex reasoning. This could make advanced AI cheaper and more reliable for many applications.