Rational Metareasoning for Large Language Models

Reasoning has emerged as a core technique for improving large language model (LLM) performance across various tasks by using additional inference-time compute. However, as LLMs scale in both size and usage, inference costs are becoming increasingly burdensome. How, then, might we optimize the cost-performance tradeoff of reasoning? This work introduces a novel approach based on computational models of metareasoning used in cognitive science, training LLMs to selectively use intermediate reasoning steps only when necessary. We first develop a reward function that incorporates the Value of Computation by penalizing unnecessary reasoning, then use this reward function with Expert Iteration to train the LLM. Compared to few-shot chain-of-thought prompting, our approach significantly reduces inference costs (38\% fewer tokens generated on average) without sacrificing task performance across diverse datasets.