按年被提到
20252025 · 5 次2026
大模型与语言模型›幻觉与语言模型局限·arXiv:2509.04664 · cs.CL

Why Language Models Hallucinate

语言模型为何会产生幻觉
Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, Edwin Zhang · 2025

幻觉源于训练与评测奖励猜测、不奖励承认不确定

开始阅读 ↗PDFalphaXiv 讨论
被推荐 · 7 人首次 2025-09-08 · 最近 2026-05-26
7不同的人
5活跃月份
8讨论串
47被引(OpenAlex)

论文认为幻觉并不神秘:在预训练中它相当于二分类错误,若无法区分错误陈述与事实,统计压力就会自然产生幻觉。后训练阶段幻觉之所以持续,是因为主流评测按对错打分,奖励猜测、惩罚表达不确定。作者主张修改那些主导排行榜、却与目标错位的基准的打分方式,而不是再新增幻觉评测。

HN 为什么提它

评论者引用它解释模型为何宁可猜也不拒答(基准奖励猜测);也有人认为标题比论文结论更耸动,并质疑其中的类比。

观点难度 3 / 5约 40 分钟经得起时间
读之前最好知道
1大语言模型预训练与后训练
2二分类与误分类误差
3基准评测(benchmark)打分
4校准(calibration)

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janalsncm · 2025-09

“Perhaps this OpenAI paper would be interesting then (published September 4th):”

推荐这篇 OpenAI 论文,讲模型不确定时会猜测
What happens when coding agents stop feeling like dialup?原评论 ↗
批评nunez · 2025-09

“This is a de facto false equivalence for two reasons.”

OpenAI’s latest research paper demonstrates that falsehoods are inevitable原评论 ↗
引用xmcqdpt2 · 2025-12

“In part because model performance is benchmarked using tests that favor giving partly correct answers as opposed to refusing to answer.”

拿它说明幻觉部分源于基准偏向给出部分正确答案而非拒答
Apple's slow AI pace becomes a strength as market grows weary of spending原评论 ↗
引用xmcqdpt2 · 2026-02

“According to that OpenAI paper, models hallucinate in part because they are optimized on benchmarks that involve guessing.”

引用它说明模型幻觉部分源于针对猜测型基准的优化
Two kinds of AI users are emerging. The gap between them is astonishing原评论 ↗
引用droidjj · 2026-05

“It's not as sensational as the headline suggests.”

提醒论文没有标题那么耸动,幻觉不可避免只针对基座模型
OpenAI admits AI hallucinations are mathematically inevitable (Sept. 2025)原评论 ↗
被提交为 story · 3 次