Meta AI推理新论文：模型记住套路，推理token砍半

Core Insights - Meta has developed a new mechanism for large language models (LLMs) that allows them to "think less and think clearer," significantly improving reasoning efficiency [1][3]. Group 1: Research Findings - The paper titled "Metacognitive Reuse: Turning Recurring LLM Reasoning Into Concise Behaviors" introduces a method where LLMs summarize their reasoning steps into concise instructions called "behaviors" [1][3]. - In mathematical reasoning tasks, the model demonstrated a reduction of up to 46% in the number of tokens required for reasoning without sacrificing accuracy [3][11]. - This mechanism is referred to as "Metacognitive Pathway," enabling models to reflect on their reasoning processes and store common strategies for future use [10][15]. Group 2: Mechanism and Implementation - The "Behavior Handbook" framework allows models to document their reasoning processes and identify common strategies, which are then named and recorded as behaviors [6][9]. - The model can call upon these behaviors when faced with similar problems, streamlining the reasoning process [10][12]. - The research outlines three modes of behavior extraction: Behavior-conditioned Inference, Behavior-guided Self-improvement, and Behavior-conditioned SFT, all leading to improved efficiency and accuracy in reasoning tasks [15]. Group 3: Experimental Results - Experiments using the R1-Llama-70B model showed that models could reduce reasoning tokens while maintaining or even improving performance [15]. - The study involved testing various student models, including Qwen3-32B and Llama-3.1-8B, with consistent results indicating a shift from slow reasoning to faster responses [15].