按年被提到
20242024 · 15 次2026
大模型与语言模型›幻觉与语言模型局限·arXiv:2401.11817 · cs.CL

Hallucination is Inevitable: An Innate Limitation of Large Language Models

幻觉不可避免:大语言模型的固有局限
Ziwei Xu, Sanjay Jain, Mohan Kankanhalli · 2024

用学习理论证明大模型的幻觉无法完全消除

开始阅读 ↗PDFalphaXiv 讨论
被推荐 · 18 人首次 2024-02-09 · 最近 2026-10-08
18不同的人
14活跃月份
24讨论串
544被引(OpenAlex)

论文把幻觉形式化为可计算的 LLM 与可计算的真值函数之间的不一致,并借助学习理论证明:LLM 无法学会所有可计算函数,因此作为通用问题求解器必然会产生幻觉。作者认为形式化世界只是真实世界的一部分,所以真实的 LLM 也不可避免地幻觉,并针对受可证明时间复杂度限制的真实 LLM 描述易幻觉的任务、做了实证验证,还讨论了现有幻觉缓解手段的机制与效果。

HN 为什么提它

评论者每逢讨论幻觉就引用它作为「幻觉不可避免」的形式化证明,有人称还没见到有力的反驳,也有人坦言只读过标题。

其他难度 4 / 5约 40 分钟经得起时间
读之前最好知道
1计算理论与可计算函数
2学习理论(对角化论证)
3大语言模型基础
4幻觉及其缓解方法

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decimalenough · 2026-06

“I'm not an expert in the field, but I've yet to see a solid rebuttal to this paper;”

推荐这篇论文,称自己还没见过对它站得住的反驳
The Permission Slip原评论 ↗
引用nyrikki · 2024-02

“But hallucinations are not avoidable either.”

用它说明幻觉同样无法避免
Alright then keep your secrets原评论 ↗
引用nyrikki · 2024-05

“Here is a formal proof.”

贴出它作为形式化证明:LLM 的幻觉无法消除
I stumbled upon LLM Kryptonite and no one wants to fix this model-breaking bug原评论 ↗
引用duskwuff · 2024-05

“My understanding was the opposite -- that the activity of a confabulating LLM is indistinguishable from one giving factually accurate responses.”

拿它支持“胡编的 LLM 与给出准确回答的 LLM 活动无法区分”的看法
Google scrambles to manually remove weird AI answers in search原评论 ↗
引用astrange · 2024-05

“This seems like it's true since LLMs are a finite size”

认为 LLM 规模有限,因此幻觉不可避免
Google scrambles to manually remove weird AI answers in search原评论 ↗
引用nyrikki · 2024-09

“I can offer cite that hallucinations are an innate property of LLMs”

拿它作为引文,证明幻觉是 LLM 的固有属性
GPTs and Hallucination原评论 ↗
看全部 20 条 →
引用az09mugen · 2024-09

“What reinforces my opinion are papers like these: [OTHER_PAPER] Hallucination is Inevitable”

引用这类论文来加强自己的观点:幻觉在 LLM 中不可避免
I no longer use AI assistant原评论 ↗
引用az09mugen · 2024-10

“some papers describe it [0] and [1] for example.”

列出它等论文,说明 LLM 幻觉不可避免的形式化论证
Announcing the Tech for Palestine Incubator原评论 ↗
引用az09mugen · 2024-10

“And researchers proven that: Hallucination is Inevitable”

称研究者已证明:无论资源多少,LLM 总会幻觉
Yann LeCun Thinks AI Is Dumber Than a Cat原评论 ↗
引用az09mugen · 2024-10

“LLM will always hallucinate. People did already some research, among which: [OTHER_PAPER] and Hallucination is Inevitable for example.”

列出它等研究,支持“LLM 总会幻觉”
Show HN: Last CLI tool you will need原评论 ↗
引用Retr0id · 2024-10

“(disclaimer, I also have not read any of these papers beyond the title!)”

只提了标题,并声明自己没读过论文内容
LLMs know more than they show: On the intrinsic representation of hallucinations原评论 ↗
引用spartanatreyu · 2024-12

“and yet, the paper that went around in March:”

用三月流传的这篇论文反驳帖子的主张:LLM 幻觉不可避免
Automated reasoning to remove LLM hallucinations原评论 ↗
引用lolinder · 2024-12

“Every article on hallucinations needs to start with this fact until we've hammered that into every "AI Engineer"'s head.”

主张每篇谈幻觉的文章都该从这个事实讲起:幻觉不可避免
AI hallucinations: Why LLMs make things up (and how to fix it)原评论 ↗
引用groby_b · 2025-04

“Still, the whole idea of probabilistic selection of tokens more or less guarantees it'll go wrong from time to time. (See e.g. Hallucination is Inevitable)”

引用论文佐证:概率式选 token 的机制决定了会时不时出错
The Policy Puppetry Attack: Novel bypass for major LLMs原评论 ↗
引用soraminazuki · 2025-09

“We get study after study and example after example showing that AI is unreliable [1]”

引用它说明幻觉是 LLM 的固有局限,AI 不可靠
MIT Study Finds AI Use Reprograms the Brain, Leading to Cognitive Decline原评论 ↗
引用solid_fuel · 2026-01

“Hallucinations are, in fact, an unavoidable property of the technology - something which is not true for people. [0]”

引用它说明幻觉是这项技术无法避免的属性,人则不同
Please Don't Say Mean Things about the AI I Just Invested a Billion Dollars In原评论 ↗
引用2ndorderthought · 2026-04

“It's literally an unavoidable fact unless we change the modelling approach.”

称除非改变建模方式,幻觉就是无法避免的事实
Which one is more important: more parameters or more computation? (2021)原评论 ↗
引用coldtea · 2026-05

“And some limitations are fundamental, and have been rigorously demonstrated, e.g.:”

用它说明有些 LLM 局限是根本性的,已被严格证明
Using "underdrawings" for accurate text and numbers原评论 ↗
引用JCTheDenthog · 2026-05

“with their inherent and unsolvable[1] hallucination problems”

给“LLM 幻觉问题固有且无解”的说法附上引用
Agents need control flow, not more prompts原评论 ↗
引用jazzypants · 2026-10

“I find hallucinations in my (mostly perfect) AI output every single day.”

以自己每天遇到的幻觉印证它:幻觉是 LLM 固有且无法避免的局限
“Math 2.0” will need to value mathematical progress more holistically原评论 ↗