按年被提到
20222023 · 17 次2026
大模型与语言模型›大模型评测与真实性·arXiv:2207.05221 · cs.CL

Language Models (Mostly) Know What They Know

语言模型(大体上)知道自己知道什么
Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, Scott Johnston, Sheer El-Showk 等 36 人 · 2022

大模型能校准地评估自己答案是否正确,并预测自己是否会答对

开始阅读 ↗PDFalphaXiv 讨论
被推荐 · 9 人首次 2022-07-13 · 最近 2025-12-09
9不同的人
10活跃月份
23讨论串
881被引(OpenAlex)

作者研究语言模型能否评估自己论断的正确性并预测能答对哪些问题。首先表明在合适的格式下,较大模型在多样的选择题和判断题上校准良好;于是让模型先给出答案,再评估该答案正确的概率 P(True),其表现、校准性和缩放趋势都令人鼓舞,参考多个自身样本后效果进一步提升。随后训练模型预测不针对具体答案的 P(IK),即「我知道答案」的概率,模型预测得不错且部分能跨任务泛化,但在新任务上校准较差。

HN 为什么提它

HN 上的人在讨论幻觉时反复把它当作证据,说明大模型大体知道自己的答案是否正确、能分辨真伪,因此有望降低幻觉。

实证研究难度 3 / 5约 90 分钟经得起时间
读之前最好知道
1大语言模型与提示格式
2模型校准(calibration)
3概率预测与评估指标
4缩放定律(scaling laws)

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xianshou · 2024-06

“Excellent background on knowledge calibration from Anthropic:”

推荐它作为 Anthropic 关于知识校准的优秀背景材料
SaySelf: Teaching LLMs to Express Confidence with Self-Reflective Rationales原评论 ↗
推荐GaggiX · 2024-10

“The linked page from OpenAI clearly show the opposite, read the "Using SimpleQA to measure the calibration of large language models" and the paper linked: Language Models (Mostly) Know What They Know”

引它反驳,说明 LLM 能知道自己不知道什么
引用brian_cloutier · 2022-08

“Given that language models seem to mostly know when they're making correct predictions [1]”

引它说明语言模型大体知道自己的预测是否正确
Chinchilla's Wild Implications原评论 ↗
引用rain_iwakura · 2022-08

“There was a recent work (Language Models (Mostly) Know What They Know, in LLM, not molecular prediction) where models could be shown to”

引它说明 LLM 能较好判断自己知道与不知道什么
AlphaFold won’t revolutionise drug discovery原评论 ↗
引用blackbear_ · 2023-05

“It has already been done: Language Models (Mostly) Know What They Know”

引它说明模型能估计自己的答案是否正确,这已经有人做过
LLM Hallucinations: why they occur and how to avoid them原评论 ↗
引用og_kalu · 2023-09

“Language Models (Mostly) Know What They Know -”

列为相关工作,关于语言模型大体知道自己知道什么
The Internal State of a Large Language Model Knows When Its Lying原评论 ↗
看全部 20 条 →
引用og_kalu · 2023-10

“Language Models (Mostly) Know What They Know - [Language Models (Mostly) Know What They Know)”

引它说明模型知道自己何时在幻觉
LLMs Confabulate Not Hallucinate原评论 ↗
引用og_kalu · 2023-10

“Language Models (Mostly) Know What They Know - [Language Models (Mostly) Know What They Know...”

引它说明模型大致知道自己何时在幻觉
LLMs Confabulate Not Hallucinate原评论 ↗
引用og_kalu · 2023-10

“Language Models (Mostly) Know What They Know - Language Models (Mostly) Know What They Know”

引它说明模型的计算能够检测幻觉
How Will AI Learn Next?原评论 ↗
引用og_kalu · 2023-10

“Language Models (Mostly) Know What They Know - Language Models (Mostly) Know What They Know”

引它说明模型的计算能够区分幻觉
LLMs can't self-correct in reasoning tasks, DeepMind study finds原评论 ↗
引用og_kalu · 2023-10

“Language Models (Mostly) Know What They Know - Language Models (Mostly) Know What They Know”

引它说明模型的计算能够区分幻觉
LLMs can't self-correct in reasoning tasks, DeepMind study finds原评论 ↗
引用og_kalu · 2023-10

“Language Models (Mostly) Know What They Know - Language Models (Mostly) Know What They Know”

引它说明 LLM 大体知道自己的答案是否为真
ChatGPT and other AI tools could disrupt scientific publishing原评论 ↗
引用og_kalu · 2023-10

“Language Models (Mostly) Know What They Know - Language Models (Mostly) Know What They Know”

引它说明 LLM 大体知道自己的答案是否为真
The Geometry of Truth: Emergent Linear Structure in How LLMs Represent Truth原评论 ↗
引用og_kalu · 2023-10

“Language Models (Mostly) Know What They Know - Language Models (Mostly) Know What They Know”

引它说明 LLM 大体知道自己的答案是否为真
"Hallucinating" AIs Sound Creative, but Let's Not Celebrate Being Wrong原评论 ↗
引用og_kalu · 2023-10

“Language Models (Mostly) Know What They Know - Language Models (Mostly) Know What They Know”

引它说明 LLM 大体知道自己的答案是否为真
The Geometry of Truth: Do LLM's Know True and False原评论 ↗
引用og_kalu · 2023-11

“Language Models (Mostly) Know What They Know - Language Models (Mostly) Know What They Know”

列为 LLM 能分辨真假的证据
Ask HN: What is your standard for judging when AGI exists?原评论 ↗
引用og_kalu · 2023-11

“Language Models (Mostly) Know What They Know - Language Models (Mostly) Know What They Know”

列为 LLM 能分辨真假的证据
Ask HN: What is your standard for judging when AGI exists?原评论 ↗
引用og_kalu · 2023-11

“Language Models (Mostly) Know What They Know - Language Models (Mostly) Know What They Know”

列为 LLM 能分辨真假的证据
Rats have an imagination, new research suggests原评论 ↗
引用famouswaffles · 2024-09

“Language Models (Mostly) Know What They Know - Language Models (Mostly) Know What They Know”

列为 LLM 对不确定性知道得比说出来的更多的证据
AI is growing faster than companies can secure it, warn industry leaders原评论 ↗
引用godelski · 2025-12

“I believe it's a reference to the paper "Language Models (Mostly) Know What They Know".”

认为标题指的就是这篇论文
Transformers know more than they can tell: Learning the Collatz sequence原评论 ↗