按年被提到
20232023 · 11 次2026
大模型与语言模型›模型表达力与计算理论·arXiv:2301.04589 · cs.CL

Memory Augmented Large Language Models are Computationally Universal

外接记忆的大语言模型具有计算通用性
Dale Schuurmans · 2023

给 LLM 配上读写记忆,就能不改权重地模拟通用图灵机

开始阅读 ↗PDFalphaXiv 讨论
被提及 · 6 人首次 2023-02-21 · 最近 2026-05-29
6不同的人
7活跃月份
12讨论串
20被引(OpenAlex)

论文证明 Transformer 类大语言模型配上外部读写记忆后具有计算通用性:只看有限长度输入的确定性语言模型等价于有限自动机,而加上读写记忆就可能处理任意大输入、模拟任意算法。作者用已有的 Flan-U-PaLM 540B 加联想式读写记忆,精确模拟了通用图灵机 U_{15,2},不需要修改模型权重,只靠设计一台由提示词编程的存储程序计算机。

HN 为什么提它

评论者反复用它说明「带外部记忆的 LLM 在理论上图灵完备、能模拟特定图灵机」,并强调其意义在于上下文学习能走多远,而非方案本身是否实用。

其他难度 4 / 5约 45 分钟经得起时间
读之前最好知道
1图灵机与图灵完备(Turing completeness)
2有限自动机
3大语言模型与提示(prompting)
4上下文学习(in-context learning)

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macawfish · 2023-02

“Memory Augmented Large Language Models are Computationally Universal”

贴出标题,说明带外部记忆的 LLM 具有计算通用性
Our Society Is Not Prepared for This Much AI Awesome原评论 ↗
引用mad0 · 2023-03

“Also see this paper: Memory Augmented Large Language Models are Computationally Universal”

建议也看这篇,讲拥有无限记忆的 LLM 能成为图灵完备
Show HN: Whatdoesthiscodedo.com – AI explanations for other people’s code原评论 ↗
引用macawfish · 2023-03

“when you add the external memory they (Flan-U-PaLM 540B to be specific) have been demonstrated to be capable of simulating a specific, well defined Turing machine without special training.”

指出加上外部记忆后,Flan-U-PaLM 540B 无需特殊训练就能模拟特定图灵机
The Unpredictable Abilities Emerging from Large AI Models原评论 ↗
引用og_kalu · 2023-04

“The significance of the paper is moreso the implications of how far ICL can take you rather than the ease/viability of the solution proposed.”

指出这篇论文的意义在于上下文学习能走多远,而非所提方案是否好用
Teaching ChatGPT to Speak My Son’s Invented Language原评论 ↗
引用og_kalu · 2023-04

“Memory augmented LLMs are computationally universal - Memory Augmented Large Language Models are Computationally Universal”

贴出它,论证带记忆增强的 LLM 具有计算通用性
Large Language Models: Scaling Laws and Emergent Properties原评论 ↗
引用spacebanana7 · 2023-05

“LLMs (at least the ones with read/write memory) can exactly simulate the execution of a universal Turing machine [1].”

引它说明带读写记忆的 LLM 能精确模拟通用图灵机的执行
Anthropic - Introducing 100K Token Context Windows, Around 75,000 Words原评论 ↗
看全部 8 条 →
引用og_kalu · 2023-10

“Transformers with memory can abstract recursively in theory. Memory Augmented Large Language Models are Computationally Universal.”

引它说明带记忆的 Transformer 理论上可以递归抽象
Ask HN: Why are current LLMs not considered AGI?原评论 ↗
引用Gomotono · 2026-05

“But there were papers in 2023 with a different approach requiring external memory Memory Augmented Large Language Models are Computationally Universal too”

补充 2023 年也有同样需要外部记忆、思路不同的论文,说明 LLM 外接记忆可具计算通用性
Claude Opus 4.8原评论 ↗
被提交为 story · 4 次