机器会不会说话里有话

「你们收信用卡吗」，四成人只回答了「收」

1979 年 Herbert Clark 让人打电话给五十家餐馆，每次都问同一句：「你们收信用卡吗？」他要看的是对方怎么答。结果分三堆：44% 只答了字面问题（「收」），38% 既答字面又把话补全（「收，Visa 和万事达都行」），16% 干脆只当请求答，直接报了卡种。同一句话，三种处理——字面那层没消失，而从字面到意图那一步也不是自动挂上的，是个可做可不做的动作。

这一步怎么做？1975 年 Grice 给的说法至今还在用：听话的人默认对方在配合这场对话，于是去问「他本可以直说，为什么偏偏这么说」，从落差里反推出没说出口的那部分。这样被反推出来的东西叫会话含义（conversational implicature）。这套推理怎么跑，「说出来的比说的少」那期讲过；这期要的是它的代价——不标价，就没法拿它比人和机器。

但「能把盐递给我吗」已经用滥了，滥到不需要现场推。1983 年 Gibbs 让受试读完这句，紧接着判断一个和「有没有能力递盐」有关的目标句：要是人真的先算字面再否掉，这个目标句该被激活、判断该变快。他没看到加速。固定成套的间接请求是被整块认出来的。

所以同一个现象有两条路：现成的句式走短路，新鲜的说法现场算。而哪些句式算「现成的」并不跨语言通用——1989 年 Blum-Kulka 等人主持的跨文化言语行为项目按直接程度给几种语言的请求排过序，阿根廷西班牙语最直接，澳大利亚英语最间接，德语比几种英语变体都直接。别把它读成礼貌排序：该项目自己的结论之一就是间接不等于礼貌。

EN In 1979 Herbert Clark had someone telephone fifty restaurants and ask the same question each time: "Do you accept credit cards?" What he wanted was the shape of the answers. They fell into three piles: 44% answered only the literal question ("yes"), 38% answered the literal question and then filled in the rest ("yes, Visa and Mastercard"), and 16% treated it purely as a request and went straight to listing the cards. One sentence, three treatments. Two things follow: the literal layer does not vanish, and the step from literal wording to intention is not automatic — it is an optional move. How is the move made? In 1975 Grice offered an account still in use: a listener assumes the speaker is cooperating, and therefore asks why this particular wording was chosen over a plainer one, recovering the unsaid part from the gap. What gets recovered that way is called conversational implicature. But "can you pass the salt" is worn smooth — worn enough that nothing needs computing. In 1983 Gibbs tested this by having people read the sentence and then immediately judge a target related to the literal question of ability. If the literal reading really came first, that target should have been primed and the judgement faster. He found no such speed-up. Frozen indirect requests are recognized whole, not routed through the literal detour. So the same phenomenon runs on two paths: ready-made forms take the shortcut; novel wordings get computed on the spot. Which forms count as ready-made, and is that constant across languages? In 1989 the Cross-Cultural Speech Act Realization Project led by Blum-Kulka and colleagues ranked requests in several languages by directness: Argentinian Spanish came out most direct, Australian English least, with German more direct than any of the English varieties. It would be a mistake to read that as a ranking of politeness — one of the project's own conclusions is that indirectness and politeness are not the same thing. The same blunt "pass me the salt" can be rude in one language's version of a situation and the default in another's.

说「有些」的时候，你在替对方清点他没说的那句话

「有些大象有鼻子。」这句听着不对劲，可它不假——所有大象都有鼻子，它当然为真。别扭的地方在于：说话人明明可以说「所有大象都有」，偏偏挑了更弱的词，于是你读出「不是全部」。这种从「他没用更强的说法」推出来的东西叫量级含义（scalar implicature）。

2001 年 Noveck 拿一批这样的句子问八到十岁的孩子和成年人，只问对不对。87% 的孩子说对，成年人里只有 41% 说对。孩子在这里反而更「逻辑」：「有些」在逻辑上和「所有」并不冲突，按字面判就该判对。是成年人多算了一步。

多算的这一步要求什么？要你手里有一张对方本来可用的说法清单，知道哪个更强哪个更弱，还确认他跳过了强的。也就是要求一个关于对方脑子的模型。

心理学里测这个模型的经典题是：给孩子看一个装着铅笔的糖果盒，问「等会儿进来的小明会以为里面装着什么」。四岁以下的孩子多半答「铅笔」——他知道的事，他默认别人也知道。这题通常四岁前后才过得去。

把两件事并排放：算出话外之意，和知道「别人不知道什么」，要的是同一样东西。所以间接说话不是修辞装饰，是一笔额外的推理开销——账单就在那儿：孩子很晚才装上它，成年人也只有四成触发。

EN "Some elephants have trunks." The sentence grates. Not because it is false — all elephants have trunks, so it is plainly true. What grates is that the speaker could have said "all elephants have trunks" and chose the weaker word instead, from which you read off "not all of them". What gets inferred from a stronger alternative having been passed over is called scalar implicature. In 2001 Noveck put sentences like this to children aged eight to ten and to adults, asking only whether each was true. 87% of the children said yes; among adults only 41% did. The children are the more logical party here: "some" does not logically conflict with "all", so judging by the literal meaning gets you the right answer. It is the adults who computed an extra step. What does that step require? It requires having a list of the wordings the speaker could have used, knowing which entries on the list are stronger and which weaker, and registering that the strong one was skipped. In other words it requires a model of the other person's mind. The classic psychological test of that model runs like this: show a child a sweets box containing pencils, then ask what a child who is about to walk in will think is inside. Children under four mostly answer "pencils" — what they know, they assume everyone knows. The task is usually passed around age four. Set the two side by side: working out what was meant, and knowing what someone else does not know, take the same machinery. Indirect speech is therefore not ornament but an extra inferential bill, and the bill shows up elsewhere: children install the step late, and even in adults scalar implicature fires only about 40% of the time. It was never automatic.

成绩单：有的项接近人，有的项塌，换个微调方式又变天

2023 年 Hu 等人编了一套题，七类语用现象各一批：善意的谎、间接请求、反讽、会话准则、隐喻、幽默、连贯。英语材料、专家编题、不给例子直接问，人和模型做同一批。当时最大的几个模型在所有项目上都好过瞎猜，有几项接近人；最差的三项是幽默、反讽、会话准则。

同年 Ruis 等人只测会话含义，把题压到只能答「是」或「否」——瞎猜就有 50%。没做过指令微调的预训练模型大约 60%，人类 86%。但同一批底座按「逐条示例」的方式微调过之后，成绩提上去一大截。这一条分量很重：如果「听懂话外之意」是规模自然带出来的本领，它不该对微调方式这么敏感——同样的参数量、同样的语料，只因为后期对齐的做法不同，成绩像换了个东西。

第三个方向的摆动来自语言：把同一批心智理论题翻成多种语言再测（XToM 那批工作），模型答「盒子里到底装着什么」各语言都稳，换成「他以为里面是什么」就往下掉，各语言掉的幅度还不一样。证据强度怎么标：三项都是单一基准、零样本提问，单拿出来都弱；但三处指向同一件事——测到的不是一项稳定本领，而是随材料、语言、训练方式大幅摆动的表现。

EN In 2023 Hu and colleagues assembled a set of items covering seven pragmatic phenomena: white lies, indirect requests, irony, conversational maxims, metaphor, humour and coherence. English materials, expert-written items, zero-shot prompting — no examples given, just the question — with humans and models answering the same set. The largest models of the day beat chance on every phenomenon and came close to human level on some; the three weakest were humour, irony and the maxims. That same year Ruis and colleagues tested conversational implicature alone, with items compressed until only "yes" or "no" was possible, making chance 50%. Pretrained models without instruction tuning scored around 60%; humans scored 86%. But the same base models, once instruction-tuned at the level of individual examples, jumped a long way up. That result carries weight. If understanding what was meant were an ability that simply arrives with scale, it should not be this sensitive to fine-tuning strategy. Same model, same parameter count, same pretraining corpus — and a different late-stage alignment recipe makes it behave like a different system on this test. There is wobble along another axis too. Translate the same battery of theory-of-mind items into several languages, as the XToM line of work did, and models answer the factual questions — what is actually in the box — stably across languages, while the belief questions — what does he think is in it — drop markedly, by differing amounts in different languages. The logic of the item is untouched; only its skin changed. How strong is this evidence? Each of these is a single benchmark on English or translated materials with zero-shot prompting, and any one of them alone counts as weak. But three independent results point the same way: what is being measured is not a stable ability, it is a performance that swings widely with materials, language and training recipe.

「它到底有没有心智理论」，被题目自己卡住了

2023 年有论文报告 GPT 系列在经典的错误信念题上达到了学龄儿童的水平。同年 Ullman 把题面改了一点点，逻辑一点不动：原题是一个贴着「巧克力」标签的袋子里装着爆米花，他把袋子改成透明的——能看见里面，标签就骗不了人。模型答错了。几种这类小改动试下来，当时几个最强的模型都掉。

2024 年 Strachan 等人做了规模大得多的一次对照：一整套成型的心理学题目（错误信念、暗示任务、怪异故事、社交失言识别），对上约一千九百名人类受试。GPT-4 在反讽、暗示、怪异故事上到了或超过人，唯一没到的是社交失言；而整体最差的 LLaMA2-70B，偏偏在社交失言上超过了人。

作者自己把原因挖出来了：原版失言题的正确答案方向是固定的，偏向那个答案的模型自动满分。他们又换种问法去问 GPT-4，发现它其实推出了「说话人并不知情」，只是不肯把结论落下来——作者称之为过度保守。

于是这里有两种完全不同的混淆。答错可能是推不出来，也可能是推出来了不肯说；答对可能是懂了，也可能是碰上一批答案方向单一的题。两种都不是模型的毛病，是题的毛病，而且都得再做一步才暴露。

争论大致三派。一派说这本来就是行为测验，人类也是靠行为被判定有心智理论的，对机器另立标准不公平；一派说人类版题目和答案大量躺在训练语料里，污染排查过不了就不算证据；还有一派主张改用要在多轮互动里更新信念的任务。三方各有最强版本，没有结论。

真正可检验的差异是这条：如果模型跑的是跟人同一套推算，它出错的方向该和人一致——人在量级含义上的失败是「退回字面」，不是乱答。所以错误模式比总分有信息，Hu 等人那篇报告的正是这个。反过来，人这侧有一条模型身上看不到的痕迹：同一个人在压力下、在不熟的外语里会整体退回字面（外语效应那期讲的就是这种退）。不付这笔代价的系统，总分打平也未必在跑同一件事。

EN In 2023 a paper reported that GPT models reached school-age performance on classic false-belief items. That same year Ullman did something very light: he altered the wording slightly while leaving the logic untouched. The original item has popcorn inside a bag labelled "chocolate". He made the bag transparent — if you can see inside, the label cannot fool you, which for a person needs no thought. The models got it wrong. He ran several such minor alterations, and the strongest models of the day all fell. In 2024 Strachan and colleagues ran a much larger comparison: a full battery of established psychological items (false belief, the hinting task, Strange Stories, faux pas recognition) against roughly 1,900 human participants. GPT-4 matched or exceeded humans on irony, hinting and the Strange Stories; the one test it did not match was faux pas. And LLaMA2-70B, the weakest model overall, beat humans on exactly that test. The authors dug out the reason: in the original faux pas items the correct answer always points the same way, so a model biased toward that answer scores perfectly by accident. They pushed one step further, rephrasing the question for GPT-4, and found it had in fact inferred that the speaker did not know — it simply would not commit to the conclusion. They call this hyperconservatism. So there are two quite different confusions in play here. A model that answers wrongly may be unable to infer, or may have inferred and declined to say. A model that answers correctly may have understood, or may have met a set of items whose answers all point one way. Neither is a flaw in the model; both are flaws in the test, and both surfaced only after extra work was done. The argument today runs roughly three ways. One camp says these are behavioural tests by construction, that humans are credited with theory of mind on behavioural grounds too, and that holding machines to a different standard is unfair. A second says the human versions of these items and their answer keys sit in the training data in bulk, and evidence that cannot clear a contamination check is not evidence. A third wants different items altogether: tasks requiring beliefs to be updated across turns of interaction, rather than a story followed by a multiple-choice question. Each camp has a strong version and the matter is unsettled. Here is a genuinely testable difference that does follow. If a model runs the same inference people run, its errors should point where human errors point — human failures on scalar implicature are retreats to the literal reading, not random answers. Error patterns therefore carry more information than totals, which is precisely what Hu and colleagues reported. On the human side there is a trace of cost that has no counterpart in a model: the same person under time pressure, or working in a language they know less well, retreats to literal readings across the board — the retreat described in the topic on the foreign-language effect. A system that pays no such bill may be scoring the same total while doing something else entirely.