按年被提到
20232023 · 19 次2026
强化学习与智能体›大模型智能体·arXiv:2303.11366 · cs.AI

Reflexion: Language Agents with Verbal Reinforcement Learning

Reflexion:用语言形式的强化学习训练语言智能体
Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, Shunyu Yao · 2023

让智能体用文字反思失败并存入记忆,不改权重也能越试越好

开始阅读 ↗PDFalphaXiv 讨论
被推荐 · 16 人首次 2023-03-25 · 最近 2026-03-02
16不同的人
9活跃月份
21讨论串
1,585被引(OpenAlex)

提出 Reflexion 框架:智能体不更新模型权重,而是对任务反馈进行语言形式的反思,把反思文本存进情景记忆缓冲区,供后续尝试参考。它在决策、编程和语言推理任务上都优于基线,例如 HumanEval 上 pass@1 达到 91%,超过当时 GPT-4 的 80%。

HN 为什么提它

HN 上的人把它作为「让 LLM 自我检查、自我纠错」的标准出处,用来说明自省循环能提升准确率、减少幻觉;也有人认为单靠循环作用有限。

方法难度 3 / 5约 50 分钟经得起时间
读之前最好知道
1大语言模型(LLM)与提示
2强化学习基础(reinforcement learning)
3思维链(chain-of-thought)提示
4语言智能体(language agent)

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
推荐avindroth · 2023-08

“Reflexion can serve as good reference here:”

推荐 Reflexion 作参考:智能体把检查器输出回馈成语言反思来自我纠正
Maccarone: AI-managed code blocks in Python原评论 ↗
推荐unoti · 2023-09

“This simple expedient increased its performance a ridiculously unexpected amount.”

称这个简单做法让表现意外地大幅提升
ChatGPT can now see, hear, and speak – openai.com原评论 ↗
引用isaacfrond · 2023-03

“the paper: Reflexion”

给出论文链接,作为超越 CoT 提示的“慢思考”思路
GPT-4 performs significantly worse on coding problems not in its training data原评论 ↗
引用Dzugaru · 2023-03

“My guess is that loops by itself (Reflexion [1], etc.) won't change much.”

提到 Reflexion,猜测单靠这类循环不会带来太大改变
Microsoft Researchers Claim GPT-4 Is Showing “Sparks” of AGI原评论 ↗
引用naasking · 2023-03

“Reflexion: an autonomous agent with dynamic memory and self-reflection, Reflexion”

介绍 Reflexion:带动态记忆和自我反思的自主智能体
ChatGPT and Code Interpreter = Magic原评论 ↗
引用sacred_numbers · 2023-03

“The Reflexion paper that came out recently shows how this kind of mistake might be overcome.”

引用它说明初稿后自我反思可提高准确率,从而克服这类错误
Show HN: Regex.ai – AI-powered regular expression generator原评论 ↗
看全部 17 条 →
引用Veen · 2023-04

“There's a lot of work happening at the moment around self-reflection and getting LLMs to identify and correct their own hallucinations and mistakes.”

说明目前有很多工作让 LLM 自我反思并纠正自己的幻觉和错误
AI won't steal your job, people leveraging AI will原评论 ↗
引用stevenhuang · 2023-04

“That part on providing the opportunity to check its answers is called reflexion and has been shown to significantly improve accuracy and reliability of LLM responses.”

称让模型检查自己的答案叫 reflexion,已被证明显著提升准确性和可靠性
Defamed by ChatGPT: Law professor libeled from whole cloth原评论 ↗
引用og_kalu · 2023-04

“Yes they do”

回答“是的”,引用它说明语言智能体能通过语言反馈反思并改进
ChemCrow: Augmenting large-language models with chemistry tools原评论 ↗
引用og_kalu · 2023-04

“and so can GPT-4 -”

附上它,称 GPT-4 也会犯类似人类的逻辑错误
Large Language Models: Scaling Laws and Emergent Properties原评论 ↗
引用stevenhuang · 2023-05

“This is called reflexion and has been found to be an effective way to improve response accuracy and usefulness (also reduces hallucinations).”

称这叫 reflexion,是提高回答准确性和有用性的有效方法,还能减少幻觉
The Leverage of LLMs for Individuals原评论 ↗
引用og_kalu · 2023-06

“Sure it can. It's not default behavior sure but it's fairly trivial to set up, just expensive.”

回应说模型当然能做到,虽非默认行为,但设置简单,只是成本高
The future of AI according to thousands of forecasters原评论 ↗
引用krychu · 2023-07

“Reflexion: Reflexion”

列出 Reflexion 的来源论文
Langchain Is Pointless原评论 ↗
引用lhl · 2023-07

“GPT-4 was tested at 80.1 pass@1 and 91 pass@1 using their Reflexion technique.”

引用结果:GPT-4 用 Reflexion 技术在 pass@1 上达到 91
引用og_kalu · 2023-07

“GPT-4 can reflect just fine”

引用它说明 GPT-4 完全能对自己的输出做反思
AI Safety and the Age of Dislightenment原评论 ↗
引用ipsum2 · 2024-10

“And so has Reflection (March 2023) Reflexion.”

引用它作为先前工作:语言智能体的语言式自我反思
Update on Reflection-70B原评论 ↗
引用inguz · 2026-03

“(not itself new, see e.g. Shinn et al Reflexion)”

指出这并非新做法,Shinn 等人已提出:智能体通过语言反思过往结果来改进
Show HN: Benchmarking the Keep memory system with LoCoMo原评论 ↗
被提交为 story · 3 次