鏈式草稿：以少寫多，加速思考

Chain of Draft: Thinking Faster by Writing Less

February 25, 2025

作者: Silei Xu, Wenhao Xie, Lingxiao Zhao, Pengcheng He

cs.AI

摘要

大型語言模型（LLMs）通過如思維鏈（Chain-of-Thought, CoT）提示等機制，在解決複雜推理任務中展現了卓越的性能，該機制強調詳細的逐步推理。然而，人類通常採用更高效的策略：起草簡潔的中間思考，僅捕捉關鍵信息。在本研究中，我們提出了一種受人類認知過程啟發的新範式——草稿鏈（Chain of Draft, CoD），其中LLMs在解決任務時生成簡約但信息豐富的中間推理輸出。通過減少冗餘並聚焦於關鍵洞察，CoD在準確性上與CoT相當或更優，同時僅使用7.6%的token，顯著降低了各種推理任務的成本和延遲。

English

Large Language Models (LLMs) have demonstrated remarkable performance in solving complex reasoning tasks through mechanisms like Chain-of-Thought (CoT) prompting, which emphasizes verbose, step-by-step reasoning. However, humans typically employ a more efficient strategy: drafting concise intermediate thoughts that capture only essential information. In this work, we propose Chain of Draft (CoD), a novel paradigm inspired by human cognitive processes, where LLMs generate minimalistic yet informative intermediate reasoning outputs while solving tasks. By reducing verbosity and focusing on critical insights, CoD matches or surpasses CoT in accuracy while using as little as only 7.6% of the tokens, significantly reducing cost and latency across various reasoning tasks.