按年被提到
20222023 · 24 次2026
大模型与语言模型›可解释性与表征研究·arXiv:2210.13382 · cs.LG

Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task

涌现的世界表征:在合成任务上训练的序列模型
Kenneth Li, Aspen K. Hopkins, David Bau, Fernanda Viégas, Hanspeter Pfister, Martin Wattenberg · 2022

在黑白棋上训练的 GPT 内部涌现出棋盘状态表征

开始阅读 ↗PDFalphaXiv 讨论
被推荐 · 20 人首次 2023-01-22 · 最近 2026-04-30
20不同的人
11活跃月份
24讨论串
104被引(OpenAlex)

作者用 GPT 变体预测黑白棋(Othello)这一简单棋类游戏的合法走法,网络事先对游戏及规则一无所知,却在内部涌现出非线性的棋盘状态表征。干预实验表明这一表征可以用来控制网络的输出,并能生成「潜在显著性图」,以人类可理解的方式帮助解释预测。

HN 为什么提它

评论者几乎都把它当作「LLM 不只是随机鹦鹉,而会在内部形成世界模型」的证据来引用,尤其是在 2023 年关于 LLM 是否理解世界的争论中。

实证研究难度 3 / 5约 40 分钟经得起时间
读之前最好知道
1GPT 与 Transformer 语言模型
2探针(probing)与可解释性
3黑白棋(Othello)规则
4干预实验

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JoshuaDavid · 2023-04

“Have a look at this research paper[1] and follow-up post[2].”

推荐这篇论文及后续文章,讲 Othello 模型形成内部棋盘表示
We need a much more sophisticated debate about AI原评论 ↗
批评pmontra · 2023-01

“I read the web page and the PDF paper quickly and it seems that there are no examples of the prompts they used.”

Do Large Language Models learn world models or just surface statistics?原评论 ↗
引用spion · 2023-02

“GPT style language models try to build a model of the world: Emergent World Representations”

引它说明 GPT 式语言模型会构建世界模型
What Is ChatGPT Doing and Why Does It Work?原评论 ↗
引用sigmoid10 · 2023-02

“Here is just one recent example: Emergent World Representations”

举它为 Transformer 从序列中形成内部世界表示的近期例子
ArxivGPT: Chrome extension that summarizes arxived research papers using ChatGPT原评论 ↗
引用kilgnad · 2023-02

“Also here's a scientific paper showing that LLMs are more then a chinese room: Emergent World Representations”

引它说明 LLM 不只是中文屋
What Is ChatGPT Doing and Why Does It Work?原评论 ↗
引用codeulike · 2023-02

“GPTs can model simpler games like Othello”

引它说明 GPT 能对 Othello 这类简单游戏建模
To Build Truly Intelligent Machines, Teach Them Cause and Effect原评论 ↗
看全部 20 条 →
引用kilgnad · 2023-02

“GPT style language models try to build a model of the world: Emergent World Representations”

引它说明 GPT 式语言模型会构建世界模型
To Build Truly Intelligent Machines, Teach Them Cause and Effect原评论 ↗
引用nayroclade · 2023-03

“See this paper Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task by Kenneth Li et al.: Emergent World Representations”

引它说明 LLM 能学会游戏规则并下棋
Bing AI can play chess原评论 ↗
引用sebzim4500 · 2023-03

“Here is an example where researchers trained a transformer to generate legal sequences of moves in the board game Othello.”

举它为训练 Transformer 生成 Othello 合法走法的例子
引用HervalFreire · 2023-03

“GPT style language models try to build a model of the world: Emergent World Representations”

引它说明 GPT 式语言模型会构建世界模型
GPT-4 Designed a Programming Language原评论 ↗
引用Herval_freire · 2023-03

“GPT style language models try to build a model of the world: Emergent World Representations”

引它说明 GPT 式模型会构建世界模型
AI-enhanced development makes me more ambitious with my projects原评论 ↗
引用sanxiyn · 2023-04

“LLMs clearly develop internal representations, this is an empirical fact. For example, one paper cited is Emergent World Representations, which trained an LLM on Othello games”

引它说明 LLM 确实形成内部表示,例如在 Othello 棋局上训练
Eight Things to Know About Large Language Models [pdf]原评论 ↗
引用patrickas · 2023-04

“This paper from late last year shows that LLMs are not "just" stochastic parrots, but they actually build an internal model of the "world" that is not programmed in, just from trying to predict the next token.”

引它说明 LLM 不只是随机鹦鹉,仅靠预测下一个 token 就形成内部世界模型
引用hackinthebochs · 2023-04

“LLMs demonstrably model their training data, which has a correspondence to the structure in the world[1][2].”

引它说明 LLM 对训练数据建模,并对应世界的结构
The LLama Effect: Leak Sparked a Series of Open Source Alternatives to ChatGPT原评论 ↗
引用og_kalu · 2023-05

“A world model is so obvious papers like these are more confirmation than surprise”

引它说明世界模型很显然,这类论文更像是印证而非惊喜
Just Calm Down About GPT-4 Already原评论 ↗
引用nopinsight · 2023-05

“Here's a research paper on this and a blog post by the lead author summarizing the results.”

贴出论文和第一作者的博客总结,说明语言模型会学出世界模型
Just Calm Down About GPT-4 Already原评论 ↗
引用Davidbrcz · 2024-05

“but it seems that models can have emerging model states of a given situation”

引它说明模型似乎能涌现出对给定情境的内部状态
Ask HN: How do you explain LLMs to those *still* unfamiliar?原评论 ↗
引用thomashop · 2024-10

“the model learns an internal representation of the board and its pieces.”

引它说明在 Othello 上训练的模型学到了棋盘与棋子的内部表示
Chain-of-thought can hurt performance on tasks where thinking makes humans worse原评论 ↗
引用Miraltar · 2024-11

“related: Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task Emergent World Representations”

列为序列模型中世界模型的相关阅读
Something weird is happening with LLMs and chess原评论 ↗
引用jeremyjh · 2026-04

“But we do not fully understand how higher-level behavior emerges - it is a subject of active research.”

举它为模型形成内部世界表示的例子,同时说明高层行为如何涌现仍在研究
Where the Goblins Came From原评论 ↗