AI 与机器学习2025arxiv.org
我们礼貌地坚持：你的 LLM 必须学会波斯 Taarof 礼仪
We Politely Insist: Your LLM Must Learn the Persian Art of Taarof
进阶 · 约 30 分钟读完前置知识：LLM 评测基础、语用学基础、波斯语文化常识
导读
这篇论文研究大语言模型能否理解波斯语中的 taarof——一种讲究客套、间接表达与婉拒的礼仪性交际规范。作者构建了 taarof 场景的评测集与标注方法，并以波斯语母语者为人类上限（平均准确率 81.8%），据此衡量模型的文化语用能力。
HN 关注点：HN 读者既好奇 LLM 能否掌握这种高度依赖文化默契的间接交际，也围绕“人类基准线为何只有 81.8%”、“人工构造数据是否合适”以及 taarof 与欺骗性委婉的边界展开讨论。
HN 评论精选
这几条评论里：质疑 1 补充 2 亲历 1
亲历波斯裔读者用亲历解释 taarof 的实际运作逻辑：好客是义务，但接受与拒绝都要拿捏分寸，否则会冒犯对方。
作为波斯人，我觉得有必要解释一下。整个中东都奉行一种慷慨好客的原则，比如阿拉伯人的 karam，甚至被视为宗教义务。所以如果你是路过某地的陌生人，和当地人聊了几句，很可能有人请你喝茶甚至吃晚饭！你该接受吗？多数西方人会想“我有时间吗”，而中东人会想“他们有时间吗”。记住，你作为客人也要表现得慷慨体贴——也就是尽量少给对方添麻烦。但客人必须权衡：拒绝招待到底是在体贴对方，还是在冒犯对方。再说 taarof。波斯人把好客推到了近乎病态的程度。这种行为常常是机械化的套路（比如一根黄瓜你至少要推辞两次才能接受），有时则离谱到极点。讲个故事：我妈和姨妈三年没见，约在咖啡馆吃午饭。两人在收银台前为谁付钱来回推让，最后吵了起来，各自坐到不同的桌子吃饭。还有一位关系很好的家庭朋友，在我妹妹毫无请求的情况下主动提出让她周末住家里，妹妹接受了，那位朋友事后却向我妈抱怨。我有个朋友去修车，师傅说小毛病免费修，朋友懂规矩……（后面省略）
fny查看英文原文
As a Persian, I feel compelled explain. The entire of the Middle East operates on a principle of magnanimous hospitality. The Arabs call this karam, for example, which is considered to even be a religious obligation. So if you're a stranger passing through a town and chat with some locals, you can expect for someone to invite you for tea and maybe even dinner! Do you take them up on the offer? Most people in the West would ask themselves "do I have the time", but a middle easterner would think "do they have the time?" Remember, you as a guest are also magnanimously hospitable--in this case, the least possible burden. Unfortunately, the guest must navigate whether they are being considerate or insulting by refusing hospitality. Now to address taarof specifically. Persian's have pathologized hosptiality to the point of psychosis. The behavior is often mechanically choreographed (e.g. you should refuse at least twice before accepting a cucumber.) In other cases, it's insane. Story time: My mom and aunt had not seen each other for 3 years and met for lunch at a cafe. They went back and forth over who would pay at the register until it erupted into a fight. They sat at separate tables for lunch. A close family friend--with no provocation from my sister--offered to house her for a weekend she was visiting. My sister accepted. That friend later complained to my mother. A friend of mine went to a mechanic who offered to fix a small issue for free. My friend knowing better paid him anyway, but the mechanic was still upset since he didn't pay enough! Edit: Just to address some of the comments. Abroad, the degree of hospitality is warped for foreigners (for what I hope are obvious reasons). Middle easterners in the West are far more inclusive.
补充人类 81.8% 的上限令人意外；taarof 式的间接表达在湾区科技圈的代码评审里同样存在，本质上是靠语用默契传递真实意图。
摘要里说“波斯语母语者构成人类上限，在 taarof 预期场景上平均准确率 81.8%”，我很惊讶人类基准这么低。taarof 的经典例子——出租车司机坚称这趟免费、但心里期望收钱——我别处也见过，这里的 taarof 对交易本身是关键性的。我猜人类只是搞错了边缘案例。顺便说，湾区科技文化里也有类似的东西。有件事让我抓狂：有人在代码评审里写“你或许可以考虑在这里用 X 数据结构”，实际意思是“你不用 X 我就不合并这段代码”。我能想象，这种 taarof 会让更字面化理解的波斯语母语者同样恼火。另外，这几乎是现实生活最接近《星际迷航》里“Darmok and Jalad at Tanagra”那种典故式沟通的情形了。
quotemstr查看英文原文
> Native Persian speakers establish the human ceiling. Native speakers achieved an average accuracy of 81.8% on taarof-expected scenarios, demonstrating high but not perfect agreement. This establishes an appropriate ceiling for model performance and further validates our annotation approach I'm surprised human benchmark is that low. The canonical example of taarof, one I've seen elsewhere, is of a taxi driver insisting that a ride is free while expecting to get paid. Taarof in this case is load-bearing for the transaction. I presume humans only get the edge cases wrong. As an aside, there are elements of this sort of thing in Bay Area tech culture too. Something that drives me nuts is someone writing on a code review "you may want to consider using the X data struct here" and meaning "I will not merge this code until you use X". I can only imagine taarof irks more literal-minded Persian speakers for the same reason. Also, this is pretty much as close as real life ever gets to a "Darmok and Jalad at Tanagra" dynamic.
质疑质疑用人工合成数据“教”模型文化规范的做法，认为自然语料才是模型习得西方社交习俗的原因。
我不太确定该怎么看这件事。第一印象是不喜欢，但也许是我的理解有偏差。ChatGPT 处理西方社交习俗得当吗？我觉得是，大概因为它在这些习俗上有海量数据，甚至有些数据明确讨论了这些习俗。人们做事、被记录、再进入 LLM。但这里说的是“人为”生成内容，让 LLM 按内容制作者期望的方式回应。也许这种做法早已存在，我也没什么立场批评？
Levitz查看英文原文
I'm not sure what to think of this, on first impression I don't like it, but maybe I have a misguided impression on how this works. Does ChatGPT properly handle western social customs? I'd say yes, and I presume that's because it has a truckload of data involving such customs and even some that explicitly talks about those customs. People do stuff, it gets recorded, then into the LLM. In this case though we are talking about "artificially" generating content such that the LLM responds how the group making the content wants. Maybe that's something that was already done and I don't really have any ground to stand on?
补充引用 19 世纪 Gobineau 对 ketman（掩饰内心）的记述，担心把这类“伪装”交际艺术教给 LLM 会带来风险。
如果用 taarof 和 ketman 训练 LLM 可能相当危险……尤其考虑到……可能产生什么后果。这种高超的欺骗艺术，或许只有俄罗斯人更胜一筹。Arthur de Gobineau《亚洲三年》（1859）写道：“波斯有一个欧洲人毫无概念、甚至难以翻译的词：ketmân。它指掩饰自己的想法、隐藏自己的观点、小心不暴露真实信念或感受。它不被视为羞耻，更不是罪行；相反，它是一种美德、义务和必需品，是生活境况强加给每个人的。践行 ketmân 不仅被允许，而且是被要求的。它在于从不让自己显得是本来面目，而总是呈现出另一副样子；在于对每个人展示最讨他喜欢的面向，采纳他的观念、品味和语言，内心却完全不同。这种持续不断的伪装练得炉火纯青，波斯人还带着一种戏弄他人的快感。他们从这种精巧的伪善中获得乐趣；这是一场游戏、一场机巧的胜利，赢家是最成功地隐藏了真相的人。”
agobineau查看英文原文
It may be quite dangerous if we train LLMs on Taarof and Ketman... especially considering... what may arise. The masterful art of deception, surpassed perhaps only by the russianes Arthur de Gobineau, Trois ans en Asie (3 years in asia) 1859: “There is in Persia a word of which Europeans have no idea, and of which it is difficult even to give them a translation: this word is ketmân. It means the dissimulation of one’s thoughts, the concealment of one’s opinions, the careful hiding of what one truly believes or feels. It is not considered a shame, still less a crime; it is, on the contrary, a virtue, a duty, and a necessity, imposed on everyone by the conditions of life. To practise ketmân is not merely permitted, it is commanded. It consists in never allowing oneself to appear as one is, but in always showing oneself otherwise; it is the art of presenting to each person the aspect that will please him most, of adopting his ideas, his tastes, his language, while inwardly remaining quite different. This perpetual exercise of disguise is carried out with a marvellous ease, and with a kind of pleasure in tricking others, which the Persians feel very keenly. They take delight in this ingenious hypocrisy; it is a game, a triumph of subtlety, in which the winner is the one who has best succeeded in hiding the truth.”
历次讨论
- 2025Hacker News 讨论帖181 分 · 122 评
同领域还值得读
DeepSeek-R1：通过强化学习激励 LLM 的推理能力
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL
研究如何用强化学习激励大语言模型的推理能力，提出 DeepSeek-R1，并探讨 RL 训练对模型推理表现的影响。
AI 与机器学习2025 · 1,351 分 · 1,056 评 · 引用 30进阶 · 约 80 分钟编辑精选1-bit LLM 时代：面向低成本计算的三值参数
The Era of 1-bit LLMs: ternary parameters for cost-effective computing
提出用三值参数 {-1, 0, 1} 替代 LLM 中的浮点权重，即所谓“1-bit LLM”。
AI 与机器学习2024 · 1,040 分 · 447 评 · 引用 19进阶 · 约 45 分钟编辑精选DeepSeek-V3.2：推进开放大语言模型的前沿
DeepSeek-v3.2: Pushing the frontier of open large language models
这是 DeepSeek-V3.2 的技术报告，延续了 DeepSeek 开源权重模型的路线，目标是让开放模型继续逼近甚至追平前沿闭源模型。
AI 与机器学习2025 · 982 分 · 465 评进阶 · 约 90 分钟编辑精选通过自我对弈与通用强化学习掌握国际象棋和将棋（AlphaZero）
Mastering Chess and Shogi by Self-Play with General Reinforcement Learning
DeepMind 的 AlphaZero 论文：用通用强化学习算法（自我博弈 + MCTS + 神经网络），从零开始仅数小时训练便在国际象棋和将棋上击败了世界顶级引擎 Stock…
AI 与机器学习2017 · 539 分 · 270 评 · 引用 13进阶 · 约 90 分钟编辑精选用强化学习发现更快的矩阵乘法算法
Discovering faster matrix multiplication algorithms with reinforcement learning
论文提出 AlphaTensor，用强化学习搜索矩阵乘法算法。
AI 与机器学习2022 · 518 分 · 111 评专业 · 约 90 分钟编辑精选