按年被提到
20222023 · 16 次2026
大模型与语言模型›大模型推理与思维链·arXiv:2201.11903 · cs.CL

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

思维链提示激发大语言模型的推理能力
Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, Denny Zhou · 2022

给出逐步推理示例,显著提升大模型推理表现

开始阅读 ↗PDFalphaXiv 讨论
被推荐 · 38 人首次 2022-04-04 · 最近 2026-09-03发表于 neural information processing systems
38不同的人
27活跃月份
41讨论串
4,505被引(OpenAlex)

提出思维链(chain-of-thought)提示:在少样本示例里给出一系列中间推理步骤,让足够大的语言模型自然涌现出复杂推理能力。在三个大语言模型上,算术、常识和符号推理任务均有提升;论文称用 8 个思维链示例提示 540B 参数模型,在 GSM8K 数学应用题上达到当时最优,超过带验证器的微调 GPT-3。

HN 为什么提它

它被反复当作原始出处引用,用来说明让模型先写出步骤就能得到更准确的答案,也被用来讨论 LLM 的推理能力、涌现现象和错误是否可纠正。

奠基之作难度 2 / 5约 40 分钟经得起时间
读之前最好知道
1大语言模型基础
2少样本提示(few-shot prompting)
3Transformer 架构
4算术与常识推理评测基准

学习路径

读这篇之前,按顺序读书架上的 3 篇 · 参考文献来自 OpenAlex,顺序和要点由 Claude Code 排出

这条路径建立 Transformer、语言模型缩放规律和少样本提示的背景,再读思维链提示如何在大模型上激发推理能力。

  1. 1
    Attention Is All You Need注意力就是你所需要的一切 · 2017

    思维链提示所用的大语言模型都基于 Transformer 架构。

    读抓自注意力和解码器堆叠结构;翻译实验略读。

  2. 2
    Scaling Laws for Neural Language Models神经语言模型的缩放定律 · 2020

    缩放定律说明模型规模带来的能力提升,对应思维链只在足够大的模型上才有效的结论。

    读抓损失随参数量、数据、算力按幂律下降的主要结论。

  3. 3
    Language Models are Few-Shot Learners语言模型是少样本学习者 · 2020

    提出少样本提示的范式,思维链提示是在示例中加入推理步骤的扩展。

    读抓上下文学习、少样本提示的设置与示例格式。

  4. 4
    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models思维链提示激发大语言模型的推理能力 · 2022 · 就是这篇

    给出逐步推理示例,显著提升大模型推理表现。

    读抓思维链示例的构造、算术、常识与符号推理实验,以及模型规模带来的涌现。

HN 书评

推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及
强烈推荐password54321 · 2026-02

“The most important paper in recent years is on CoT [Chain-of-Thought Prompting Elicits Reasoning in Large Language Models and I'm not even sure what comes close.”

称 CoT 是近年最重要的论文,想不出有什么能相比
Andrej Karpathy talks about "Claws"原评论 ↗
推荐xianshou · 2023-04

“Better read: the original paper demonstrating that prompting chain-of-thought reasoning elicits superior results.”

推荐更值得读的原始论文:思维链提示能得到更好的结果
Input to GPT Matters a Lot原评论 ↗
推荐KTibow · 2025-05

“I can point you to the original chain of thought paper [0]”

指向原始 chain-of-thought 论文,作为背景资料
I used o3 to find a remote zeroday in the Linux SMB implementation原评论 ↗
批评cubefox · 2025-10

“But clearly the technique was not invented by these authors.”

Ortega hypothesis原评论 ↗
引用visarga · 2022-04

“Interesting that they used Chain of Thought Prompting[1] for improved reasoning so soon after its publication.”

感叹 PaLM 在该方法发表后这么快就用上了 Chain of Thought 提示
Pathways Language Model (PaLM): Scaling to 540B parameters原评论 ↗
引用DavidSJ · 2022-05

“But they can be told to, and then the same network becomes much smarter: Chain-of-Thought Prompting Elicits Reasoning in Large Language Models”

引用它说明让网络逐步推理后,同一个网络会聪明得多
Limitations of DALL-E原评论 ↗
看全部 20 条 →
引用skybrian · 2022-05

“You're right for GPT 3, but it's an example of chain of thought reasoning, which seems to be a new area of research [1] and might get integrated into newer versions:”

把 chain of thought 作为新兴研究方向,可能被整合进新版本
Teaching GPT-3 to reverse words原评论 ↗
引用cmorez · 2022-07

“Or force the model to produce so called "chains-of-thought"[1] to make sure the model never leaves an instructional context...”

提出可用 chain-of-thought 迫使模型不脱离指令语境
Google fires engineer who called its AI sentient原评论 ↗
引用yababa_y · 2022-07

“no room for chain-of-thought, nothing in the prompt that it could append something to.”

借 chain-of-thought 说明提示里没有留出空间,模型无从续写
Am I doing it wrong? (OpenAI test)原评论 ↗
引用_0ffh · 2022-10

“Btw, here's an example of how even a very simple zero-shot/prompt-engineering attempt to introduce a bit of system 2 reasoning into a language model can improve results.”

举例说明简单的零样本提示也能给语言模型引入一点 system 2 式推理,改善结果
AI language models are struggling to “get” math原评论 ↗
引用tuxracer · 2022-12

“If you ask it to write out the steps ahead of time you'll often get a more accurate answer [1]”

引用它说明让模型先写出步骤,往往答案更准确
A new chat feature has been released by You Search原评论 ↗
引用eternalban · 2023-01

“Using chain of thought [slightly modified] example from Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (ref'd in OP)”

用了这篇论文(原帖引用)中稍作修改的思维链示例
ChatGPT scores 80% on SAT reading/writing with collective chain of thought原评论 ↗
引用danielbln · 2023-02

“factuality can effectively be solved by chain-of-thought prompting[0]”

引用它主张事实性问题可通过思维链提示有效解决
Our Society Is Not Prepared for This Much AI Awesome原评论 ↗
引用williamcotton · 2023-03

“Chain-of-Thought Prompting Elicits Reasoning in Large Language Models: Chain-of-Thought Prompting Elicits Reasoning in Large Language Models”

贴出思维链论文,作为 LLM 错误可以修复的证据
Oceania Always at War with Eastasia: Generative AI and Knowledge Pollution原评论 ↗
引用nl · 2023-03

“It doesn't actually. See Chain-of-Thought Prompting Elicits Reasoning in Large Language Models”

反驳说 few-shot 示例源自 GPT-3 论文,并非 CoT
We gave GPT-3.5 tools to run, write, commit, and deploy code原评论 ↗
引用Dzugaru · 2023-03

“The problem is known for a while, for example this article [0] mentions it (as Chain of Thought reasoning).”

指出这个回溯问题早已为人所知,该文称之为 Chain of Thought 推理
GPT-4 performs significantly worse on coding problems not in its training data原评论 ↗
引用vsareto · 2023-03

“You'll be wanting to find emergent stuff like this and not just crafting prompts to get the right response: Chain-of-Thought Prompting Elicits Reasoning in Large Language Models”

建议去找这类涌现现象,而不只是精心设计提示来得到正确回答
$335,000 Pay for ‘AI Whisperer’ Jobs Appears in Red-Hot Market原评论 ↗
引用BiasRegularizer · 2023-04

“One additional supporting evidence on the finite step reasoning hypothesis is that "train of thought" improves the performance of LLMs[1]”

引用它作为有限步推理假说的又一证据:思维链能提升 LLM 表现
Input to GPT Matters a Lot原评论 ↗
引用sp332 · 2023-04

“can really help some of the larger GPT models. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models”

指出思维链提示确实能帮到一些较大的 GPT 模型
A guide to prompting AI (for what it is worth)原评论 ↗
引用cosmojg · 2023-05

“Just speculating, but this looks like precisely the sort of result one might expect from fine-tuning models on chain-of-thought-prompted interactions like those described by Wei et al. in 2022[1].”

推测这个结果正是在思维链提示交互上微调模型后会出现的,如 Wei 等人所述
LLMs are not greater than the sum of their parts: Stanford researchers原评论 ↗