Meta AI Proposes Metacognitive Reuse: Turning LLM Chains-of-Thought into a Procedural Handbook that Cuts Tokens by 46%

Meta researchers have developed "Metacognitive Reuse," a novel approach that compresses common reasoning patterns into concise, named procedures called "behaviors," which are stored in a searchable handbook. This method enables large language models (LLMs) to reuse these behaviors during inference, significantly reducing reasoning token usageup to 46% on the MATH datasetwhile maintaining or improving accuracy, and achieving up to 10% gains in self-improvement scenarios like AIME, all without altering model weights. The process involves a reflection pipeline where an LLM identifies recurring procedural steps from prior problem traces,