按年被提到
20222023 · 16 次2026
大模型与语言模型›大模型评测与真实性·arXiv:2205.14334 · cs.CL

Teaching Models to Express Their Uncertainty in Words

教模型用自然语言表达自身的不确定性
Stephanie Lin, Jacob Hilton, Owain Evans · 2022

GPT-3 可用文字说出自己答案的置信度,且校准良好

开始阅读 ↗PDFalphaXiv 讨论
被提及 · 4 人首次 2022-05-31 · 最近 2024-10-30
4不同的人
6活跃月份
18讨论串
184被引(OpenAlex)

论文表明 GPT-3 可以学会用自然语言(如“90% 置信度”“高置信度”)表达对自己答案的把握,而不依赖 logits,这些置信度与实际正确率大体校准。作者提出 CalibratedMath 任务集,对比了“口头概率”与从 logits 提取的不确定性,两者在分布偏移下都能保持一定校准。结果还显示这种能力依赖预训练中已形成的、与认知不确定性相关的潜在表示。

HN 为什么提它

讨论 LLM 是否知道自己在幻觉、能否分辨真假时,评论者反复把它作为模型能表达校准后不确定性的证据。

实证研究难度 3 / 5约 45 分钟经得起时间
读之前最好知道
1大语言模型微调(fine-tuning)
2模型校准(calibration)
3概率与置信度
4分布偏移(distribution shift)

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gwern · 2022-05

“By the way, did you know that this 'big blob of inanimate knowledge' also knows what it doesn't know and can give well-calibrated predictions of accuracy Teaching Models to Express Their Uncertainty in Words ?”

引用它说明模型也知道自己不知道什么,能对准确率给出校准良好的预测
The Scaling Hypothesis (2021)原评论 ↗
引用og_kalu · 2023-09

“Teaching Models to Express Their Uncertainty in Words -”

把这篇列为相关工作,关于教模型用语言表达不确定性
The Internal State of a Large Language Model Knows When Its Lying原评论 ↗
引用og_kalu · 2023-10

“Teaching Models to Express Their Uncertainty in Words - [Teaching Models to Express Their Uncertainty in Words)”

引用它作证据:模型知道自己何时在产生幻觉
LLMs Confabulate Not Hallucinate原评论 ↗
引用og_kalu · 2023-10

“Teaching Models to Express Their Uncertainty in Words - [Teaching Models to Express Their Uncertainty in Words...”

引用它作证据:模型大致知道自己何时在产生幻觉
LLMs Confabulate Not Hallucinate原评论 ↗
引用og_kalu · 2023-10

“Teaching Models to Express Their Uncertainty in Words - Teaching Models to Express Their Uncertainty in Words”

引用它作证据:模型的计算能检测出幻觉
How Will AI Learn Next?原评论 ↗
引用og_kalu · 2023-10

“Teaching Models to Express Their Uncertainty in Words - Teaching Models to Express Their Uncertainty in Words”

引用它作证据:模型的计算能区分幻觉
LLMs can't self-correct in reasoning tasks, DeepMind study finds原评论 ↗
看全部 20 条 →
引用og_kalu · 2023-10

“Teaching Models to Express Their Uncertainty in Words - Teaching Models to Express Their Uncertainty in Words”

引用它作证据:模型的计算能区分幻觉
LLMs can't self-correct in reasoning tasks, DeepMind study finds原评论 ↗
引用og_kalu · 2023-10

“Teaching Models to Express Their Uncertainty in Words - Teaching Models to Express Their Uncertainty in Words”

引用它说明模型如何用语言表达经过校准的不确定性
ChatGPT and other AI tools could disrupt scientific publishing原评论 ↗
引用og_kalu · 2023-10

“Teaching Models to Express Their Uncertainty in Words - Teaching Models to Express Their Uncertainty in Words”

引用它说明模型如何用语言表达经过校准的不确定性
The Geometry of Truth: Emergent Linear Structure in How LLMs Represent Truth原评论 ↗
引用og_kalu · 2023-10

“Teaching Models to Express Their Uncertainty in Words - Teaching Models to Express Their Uncertainty in Words”

引用它说明模型如何用语言表达经过校准的不确定性
"Hallucinating" AIs Sound Creative, but Let's Not Celebrate Being Wrong原评论 ↗
引用og_kalu · 2023-10

“Teaching Models to Express Their Uncertainty in Words - Teaching Models to Express Their Uncertainty in Words”

引用它说明模型如何用语言表达经过校准的不确定性
The Geometry of Truth: Do LLM's Know True and False原评论 ↗
引用og_kalu · 2023-11

“Teaching Models to Express Their Uncertainty in Words - Teaching Models to Express Their Uncertainty in Words”

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

“Teaching Models to Express Their Uncertainty in Words - Teaching Models to Express Their Uncertainty in Words”

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

“Teaching Models to Express Their Uncertainty in Words - Teaching Models to Express Their Uncertainty in Words”

列为 LLM 能分辨真假的证据
Rats have an imagination, new research suggests原评论 ↗
引用og_kalu · 2023-11

“Teaching Models to Express Their Uncertainty in Words - Teaching Models to Express Their Uncertainty in Words”

列为 LLM 能分辨真假的证据
Role play with large language models – Nature原评论 ↗
引用og_kalu · 2023-11

“Teaching Models to Express Their Uncertainty in Words - Teaching Models to Express Their Uncertainty in Words”

列为 LLM 能分辨真假的证据
Language models and linguistic theories beyond words原评论 ↗
引用og_kalu · 2023-11

“Teaching Models to Express Their Uncertainty in Words - Teaching Models to Express Their Uncertainty in Words”

列为 LLM 能分辨真假的证据
I disagree with Geoff Hinton regarding "glorified autocomplete"原评论 ↗
引用famouswaffles · 2024-09

“Teaching Models to Express Their Uncertainty in Words - Teaching Models to Express Their Uncertainty in Words”

引用它作证据:LLM 对不确定性知道的比它说出来的多
AI is growing faster than companies can secure it, warn industry leaders原评论 ↗
引用famouswaffles · 2024-09

“Teaching Models to Express Their Uncertainty in Words - Teaching Models to Express Their Uncertainty in Words”

列为 LLM 内部知道的比说出来的更多的证据
GPTs and Hallucination原评论 ↗
引用famouswaffles · 2024-10

“Teaching Models to Express Their Uncertainty in Words - Teaching Models to Express Their Uncertainty in Words”

列为 LLM 不确定性方面的相关工作
LLMs know more than they show: On the intrinsic representation of hallucinations原评论 ↗