“Have a look at this research paper[1] and follow-up post[2].”
Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task
在黑白棋上训练的 GPT 内部涌现出棋盘状态表征
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。
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。
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“I read the web page and the PDF paper quickly and it seems that there are no examples of the prompts they used.”
“GPT style language models try to build a model of the world: Emergent World Representations”
“Here is just one recent example: Emergent World Representations”
“Also here's a scientific paper showing that LLMs are more then a chinese room: Emergent World Representations”
“GPTs can model simpler games like Othello”
看全部 20 条 →
“GPT style language models try to build a model of the world: Emergent World Representations”
“See this paper Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task by Kenneth Li et al.: Emergent World Representations”
“Here is an example where researchers trained a transformer to generate legal sequences of moves in the board game Othello.”
“GPT style language models try to build a model of the world: Emergent World Representations”
“GPT style language models try to build a model of the world: Emergent World Representations”
“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”
“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.”
“LLMs demonstrably model their training data, which has a correspondence to the structure in the world[1][2].”
“A world model is so obvious papers like these are more confirmation than surprise”
“Here's a research paper on this and a blog post by the lead author summarizing the results.”
“but it seems that models can have emerging model states of a given situation”
“the model learns an internal representation of the board and its pieces.”
“related: Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task Emergent World Representations”
“But we do not fully understand how higher-level behavior emerges - it is a subject of active research.”