We explore how generating a chain of thought -- a series of intermediate reasoning steps -- significantly improves the ability of large language models to perform complex reasoning. In particular, we show how such reasoning abilities emerge naturally in sufficiently large language models via a simple method called chain of thought prompting, where a few chain of thought demonstrations are provided as exemplars in prompting. Experiments on three large language models show that chain of thought prompting improves performance on a range of arithmetic, commonsense, and symbolic reasoning tasks. The empirical gains can be striking. For instance, prompting a 540B-parameter language model with just eight chain of thought exemplars achieves state of the art accuracy on the GSM8K benchmark of math word problems, surpassing even finetuned GPT-3 with a verifier.
核心贡献 · Key contributions
提出思维链提示法，一种通过包含中间推理步骤的少样本示例激发大语言模型多步推理的简单方法。 Proposes chain-of-thought prompting, a simple method that elicits multi-step reasoning in large language models via few-shot exemplars with intermediate reasoning steps.
证明思维链提示法在算术、常识和符号推理任务上显著提升性能，在 GSM8K 上达到最先进水平。 Demonstrates that chain-of-thought prompting significantly improves performance on arithmetic, commonsense, and symbolic reasoning tasks, achieving state-of-the-art on GSM8K.
表明思维链推理是模型规模的涌现能力，仅对约 1000 亿参数及以上的模型有效。 Shows that chain-of-thought reasoning is an emergent ability of model scale, only benefiting models with ~100B parameters or more.
通过消融实验分离了仅方程、可变算力和答案后思维链的效果，确认了顺序自然语言推理的重要性。 Provides ablation studies isolating the effects of equation-only, variable compute, and chain-of-thought after answer, confirming the importance of sequential natural language reasoning.
证明了思维链提示法对不同标注者、示例和语言模型的鲁棒性。 Demonstrates robustness of chain-of-thought prompting to different annotators, exemplars, and language models.
表明思维链提示法促进了长度泛化，能处理比示例中步骤更多的分布外输入。 Shows that chain-of-thought prompting facilitates length generalization to out-of-distribution inputs with more steps than seen in exemplars.
局限 · Limitations
未解决神经网络是否真正推理的问题；思维链模仿人类思维过程但不保证推理。 Does not address whether the neural network is actually reasoning; chain-of-thought emulates human thought processes but does not guarantee reasoning.
手动标注思维链示例在少样本提示中成本高，在微调中若无合成数据则难以承受。 Manual annotation of chain-of-thought exemplars is costly for few-shot prompting, and prohibitive for finetuning without synthetic data.
不保证推理路径正确；生成的思维链可能导致正确或错误答案。 No guarantee of correct reasoning paths; generated chains can lead to both correct and incorrect answers.
思维链推理仅在大型模型规模下涌现，使得在实际应用中部署成本高昂。 Chain-of-thought reasoning only emerges at large model scales, making it costly to serve in real-world applications.
评估仅限于算术、常识和符号推理；在其他领域的适用性尚未探索。 Limited evaluation to arithmetic, commonsense, and symbolic reasoning; applicability to other domains remains unexplored.