按年被提到
20222023 · 7 次2026
大模型与语言模型›大模型推理与思维链·arXiv:2205.11916 · cs.CL

Large Language Models are Zero-Shot Reasoners

大语言模型是零样本推理者
Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, Yusuke Iwasawa · 2022

只加一句“让我们一步步思考”,LLM 的零样本推理大幅提升

开始阅读 ↗PDFalphaXiv 讨论
被提及 · 14 人首次 2022-05-25 · 最近 2026-01-20
14不同的人
14活跃月份
15讨论串
2,415被引(OpenAlex)

论文表明,只要在每个回答前加上“Let's think step by step”,LLM 就是不错的零样本推理者。这个称为 Zero-shot-CoT 的单一提示模板,无需手写少样本示例,就在算术、符号推理和其他逻辑推理基准上显著超过普通零样本表现,例如用 text-davinci-002 把 MultiArith 的准确率从 17.7% 提到 78.7%,GSM8K 从 10.4% 提到 40.7%,540B 的 PaLM 上也有类似幅度的提升。作者希望它成为这些推理基准上最简单而强的零样本基线。

HN 为什么提它

评论者把它当作“Let's think step by step”提示技巧和零样本思维链的出处,用来说明 LLM 的推理表现不是单靠记忆或模式匹配。

方法难度 2 / 5约 30 分钟经得起时间
读之前最好知道
1大语言模型与提示(prompting)
2思维链提示(chain-of-thought prompting)
3少样本学习(few-shot learning)
4算术与符号推理基准

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
引用Isinlor · 2022-06

“But in fact just prompting a model like GPT-3 or Gopher with "let's think step by step" and letting it process in context is already significantly improving performance on wide variety of reasoning tasks [0].”

引用它说明只需提示“let's think step by step”就能显著提升多种推理任务表现
LaMDA is not sentient原评论 ↗
引用simonw · 2023-01

“Even the classic "let's think about this step by step" paper came out less than a year ago: Large Language Models are Zero-Shot Reasoners - May 2022.”

指出经典的“逐步思考”论文发表还不到一年,2022 年 5 月
ChatGPT Cheat Sheet原评论 ↗
引用vsareto · 2023-02

“difficult system-2 tasks that do not follow the standard scaling laws for LLMs Large Language Models are Zero-Shot Reasoners”

引用它作为背景:不遵循标准 scaling laws 的困难 system-2 任务
Tech’s hottest new job: AI whisperer. No coding required原评论 ↗
引用andyk · 2023-03

“[1] Large Language Models are Zero-Shot Reasoners - Large Language Models are Zero-Shot Reasoners”

列为参考:它提出用逐步提示激发 LLM 推理
Show HN: Recursive LLM Prompts原评论 ↗
引用convexfunction · 2023-05

“"Let's think step by step:" at the end of your prompt is probably the best-known one: Large Language Models are Zero-Shot Reasoners”

指出提示末尾加“Let's think step by step”这一最知名技巧出自该论文
Re-implementing LangChain in 100 lines of code原评论 ↗
引用TeMPOraL · 2023-07

“the title of which is, I believe, the source of the”

认为这篇论文的标题是“zero-shot”这一说法的来源,更相关
What we learned from using GPT for 500k+ classifications原评论 ↗
看全部 13 条 →
引用joshka · 2023-10

“1: Large Language Models are Zero-Shot Reasoners”

引用它作为“逐步思考”提示实现零样本推理的出处
Think Before You Speak: Training Language Models with Pause Tokens原评论 ↗
引用kromem · 2023-11

“even something as simple as a zero shot prompt followed by "let's think step by step" has significant improvement: Large Language Models are Zero-Shot Reasoners”

引用它:零样本提示后加“let's think step by step”就能显著改善
Compare how GPT-2, 3, 3.5 and 4 answer the same questions原评论 ↗
引用jdietrich · 2024-07

“their performance at zero-shot problems is simply impossible through pure memorisation or pattern-matching”

引用它论证:零样本问题上的表现不可能靠单纯记忆或模式匹配做到
Large models of what? Mistaking engineering achievements for linguistic agency原评论 ↗
引用potatoman22 · 2024-09

“to elicit a chain-of-thought response. See: Large Language Models are Zero-Shot Reasoners”

引用它说明加上逐步思考能激发 chain-of-thought 式回答
Llama 3.2: Revolutionizing edge AI and vision with open, customizable models原评论 ↗
引用TeMPOraL · 2025-03

“[2] - Large Language Models are Zero-Shot Reasoners”

列为参考:2022 年“let's think step by step”提示结果的出处
ChatGPT can't kill anything worth preserving原评论 ↗
引用LordDragonfang · 2025-06

“This technique and name stem from the early observation that LLMs do better when explicitly told to”

说明这一技术及其名称源于早期观察:明确要求时 LLM 表现更好
Magistral — the first reasoning model by Mistral AI原评论 ↗
引用CamperBob2 · 2026-01

“Since roughly 2022, they understand that they are working with systems capable of reasoning: Large Language Models are Zero-Shot Reasoners”

引用它作为 LLM 具备推理能力的证据,大约从 2022 年起人们就认识到这点
LLMs are a 400-year-long confidence trick原评论 ↗
被提交为 story · 2 次