Is AI Reasoning Right for the Wrong Reasons?

I ’ll just say it: What the hell is going on with AI “reasoning”?

我就说一句：AI“推理”到底是怎么回事？

Sorry for the air quotes.

抱歉，空气报价。

That punctuational side-eye was more common in 2024, when the specially trained cousins of LLMs now known as “large reasoning models,” or LRMs, were still new.

这种标点符号的侧视在 2024 年更为常见，当时经过专门训练的法学硕士近亲（现在被称为“大型推理模型”或 LRM）仍然是新生事物。

Nowadays it may seem downright churlish, though, given that a “general-purpose reasoning model” from OpenAI solved a famous open mathematical research problem in one shot in May 2026.

不过，考虑到 OpenAI 的“通用推理模型”在 2026 年 5 月一次性解决了一个著名的开放数学研究问题，现在看来这似乎是彻头彻尾的无礼。

Still, I’m not sure how else to acknowledge my intellectual whiplash over the scientific interpretation of what these AI systems are actually doing.

尽管如此，我还是不知道如何才能承认我对这些人工智能系统实际行为的科学解释的智力上的鞭笞。

Are we at the point where our technology is actually “reasoning,” or should that word still live within its scare quotes?

我们的技术是否真正处于“推理”阶段，或者这个词是否应该仍然存在于其吓人的引号中？

Reasoning comes in many technically defined forms , but the basic procedure is easily recognizable: arriving at a sound conclusion by linking together intermediate steps that logically follow from each other.

推理有许多技术上定义的形式，但基本程序很容易识别：通过将逻辑上相互遵循的中间步骤链接在一起，得出合理的结论。

We do this with thoughts; LRMs use so-called chains of thought, a term of art for the streams of synthetic text that the models emit before arriving at an answer to a complex query.

我们用思想来做到这一点； LRM 使用所谓的思想链，这是一个术语，指模型在得出复杂查询的答案之前发出的合成文本流。

One minute, the idea that AI could reason via these chains was being prominently and credibly critiqued (by a team of researchers from Apple) as an “ Illusion of Thinking ” subject to “complete accuracy collapse” under surprisingly simple conditions.

前一分钟，人工智能可以通过这些链条进行推理的想法受到了（来自苹果公司的一组研究人员）的显着而可信的批评，称其为“思维的幻觉”，在令人惊讶的简单条件下会“完全准确崩溃”。

The next minute, LRMs were bagging gold medals at the International Mathematical Olympiad, a feat so challenging that “even very successful mathematicians and scientists may well highlight [it] on their CVs all their lives,” as the scientist and AI critic Gary Marcus and Ernest Davis wrote in 2025.

下一分钟，LRM 就在国际数学奥林匹克竞赛中夺得了金牌，这一壮举是如此具有挑战性，以至于“即使是非常成功的数学家和科学家也可能会在他们的简历中终生强调它”，正如科学家兼人工智能评论家加里·马库斯和欧内斯特·戴维斯在 2025 年所写的那样。

If that’s not a sign of “real” reasoning, what is?

如果这不是“真正”推理的标志，那什么才是？

But wait — soon after, more research, from the Santa Fe Institute, showed that LRMs can crush even carefully designed benchmarks for reasoning (like a collection of analogy-like visual puzzles) using mere “ surface-level ‘shortcuts.’ ” What they were doing looked less like generalizable reasoning than just gaming the system.

但等等——不久之后，来自圣达菲研究所的更多研究表明，LRM 甚至可以仅使用“表面级‘捷径’”来粉碎精心设计的推理基准（例如一系列类似类比的视觉谜题）。它们所做的事情看起来不太像一般推理，而只是玩弄系统。

Then, as if on cue, another “hold my beer” moment: Google DeepMind and the mathematician Terence Tao ( the GOAT!

然后，仿佛在暗示，另一个“别喝啤酒”的时刻：谷歌 DeepMind 和数学家陶哲轩（山羊！

) used AI to rediscover or improve the solutions to 67 problems “spanning mathematical analysis, combinatorics, geometry, and number theory.” Deal with it, haters!

）使用人工智能重新发现或改进了“涵盖数学分析、组合学、几何和数论”的 67 个问题的解决方案。对付它吧，仇恨者！

I n philosophy, “qualia” refers to the subjective qualities of our experience: what it’s like for Alice to see blue or for Bob to feel delighted.

在哲学中，“感受性”指的是我们体验的主观品质：爱丽丝看到蓝色或鲍勃感到高兴是什么感觉。

Qualia are “the ways things seem to us,” as the late philosopher Daniel Dennett put it.

正如已故哲学家丹尼尔·丹尼特所说，感受性是“事物在我们看来的方式”。

In these essays, our columnists follow their curiosity, and explore important but not necessarily answerable scientific questions.

在这些文章中，我们的专栏作家追随他们的好奇心，探索重要但不一定可以回答的科学问题。

What about additional evidence that LRMs can’t reason reliably , even when they possess the necessary algorithm and computational budget to do so, and suffer from a list of scientifically documented failure states long enough to use as a Slip ’N Slide?

还有其他证据表明 LRM 无法可靠地推理，即使它们拥有必要的算法和计算预算，并且遭受了一系列科学记录的故障状态足够长的时间以用作 Slip 'N Slide？

Whatever — I guess that’s just “jagged intelligence” for you (AI-speak for “when it works, it works”).

不管怎样——我想这对你来说只是“锯齿状智能”（人工智能的意思是“当它起作用时，它就起作用”）。

And so it went from late 2025 into 2026.

就这样从2025年底一直到2026年。

I’ve been a science journalist for 20 years and an AI journalist for half of that, so I know better than to expect tidy consistency out of rapidly advancing research.

我已经做了 20 年的科学记者，其中一半是人工智能记者，所以我知道不能指望快速推进的研究能够保持整齐划一的一致性。

But even for me, this back-and-forth has been a bit much.

但即使对我来说，这种来来回回也有点过分了。

To quote Al Pacino in The Insider , “I’m getting two things: pissed off, and curious.” I don’t believe there’s fraud to be found here.

引用阿尔·帕西诺在《局内人》中的话：“我得到了两件事：生气和好奇。”我不相信这里有欺诈行为。

I just want to know which way is up.

我只是想知道哪条路是向上的。

Can AI reasoning somehow be both BS and not at the same time?

人工智能推理是否可以同时是废话？

And if so, how on Earth does that work?

如果是这样，那到底是如何运作的呢？

I knew just who to call first.

我知道该先给谁打电话。

Melanie Mitchell’s career in AI stretches back to the 1980s, but lately she’s earned a reputation as an au courant AI truth teller, penning lucid explainers for Science and her widely read newsletter , as well as conducting research at the Santa Fe Institute.

Melanie Mitchell 在人工智能领域的职业生涯可以追溯到 20 世纪 80 年代，但最近她赢得了“一位真正的人工智能说真话者”的声誉，为《科学》杂志撰写了清晰的解释文章和她广为阅读的时事通讯，并在圣达菲研究所开展研究。

(The study about “surface-level ‘shortcuts’” is hers.) When I asked her what we actually know about AI reasoning, her answer was brief enough to fit on an index card.

（关于“表面‘捷径’”的研究是她的。）当我问她我们对人工智能推理到底了解多少时，她的回答足够简短，可以放在索引卡上。

“Number one: It works.

“第一：它有效。

It improves things,” she said, referring to LRMs’ superior accuracy on reasoning tasks compared to LLMs.

它改善了事情，”她说，指的是与法学硕士相比，LRM 在推理任务上具有更高的准确性。

“Number two: The actual text that’s generated” — i.e., the chain of thought that every LRM is trained to produce to improve its performance — “isn’t necessarily faithful to what’s going on [inside the model].

“第二：生成的实际文本”——即每个 LRM 被训练以提高其性能的思想链——“不一定忠实于[模型内部]正在发生的事情。

And number three: A lot of that text isn’t even useful.

第三：很多文字甚至没有用。

You can actually take it out.”

其实你可以把它拿出来。”

Let’s unpack numbers two and three, because that’s where the superposition of “BS and not” actually lives.

让我们来解开第二和第三个数字，因为这就是“BS 与 Not”的叠加实际上存在的地方。

Chains of thought were half-discovered, half-devised in 2022 as a prompting hack for LLMs: Provide them with examples of written-out reasoning (or, famously, just ask them to “ think step by step ”), and they’ll suddenly give less boneheaded answers to simple logic and math problems.

2022 年，思想链被半发现、半设计，作为对法学硕士的一种提示：为他们提供书面推理的例子（或者，众所周知，只是要求他们“一步一步思考”），他们会突然对简单的逻辑和数学问题给出不那么愚蠢的答案。

LRMs, starting with OpenAI’s o1 model in 2024, are trained to automate this trick by generating such prompts — also called reasoning traces or thinking tokens — and then feeding them back to themselves.

LRM 从 2024 年 OpenAI 的 o1 模型开始，经过训练，可以通过生成此类提示（也称为推理痕迹或思维标记）来自动化此技巧，然后将其反馈给自身。

Because LRMs are essentially just language models, those extra bits of text create what looks convincingly like a paper trail of the model’s “thought process.”

因为 LRM 本质上只是语言模型，所以这些额外的文本位创建了看起来令人信服的模型“思维过程”的纸质记录。

Except it’s not that simple.

但事情没那么简单。

A growing body of academic and industry research has cast doubt on whether these “intermediate tokens” are a faithful representation of an LRM’s inner workings.

越来越多的学术和行业研究机构对这些“中间代币”是否忠实代表了 LRM 的内部运作提出了质疑。

Instead of being auditable receipts or accurate reports, they can appear more like what the Arizona State University researcher Subbarao Kambhampati calls “mumblings” — bits of language, yes, but ones whose meaning may be entirely incidental to any reasoning that might have occurred.

它们不是可审计的收据或准确的报告，而是更像亚利桑那州立大学研究员 Subbarao Kambhampati 所说的“咕哝”——是的，是一些语言，但其含义可能完全与可能发生的任何推理无关。

Kambhampati’s lab showed in 2025 that fully replacing a model’s correct “traces” with incorrect or irrelevant ones didn’t degrade its performance on a formal reasoning task.

Kambhampati 的实验室在 2025 年表明，用不正确或不相关的“痕迹”完全替换模型的正确“痕迹”并不会降低其在正式推理任务中的性能。

Meanwhile, training the model only on correct trace data still led it to occasionally generate invalid records of its reasoning — even when it produced a correct solution to the original problem it was given.

与此同时，仅使用正确的跟踪数据训练模型仍然会导致其偶尔生成无效的推理记录——即使它为所给出的原始问题提供了正确的解决方案。

A 2024 paper from researchers at New York University showed that “meaningless filler tokens” — literally, strings of dots — could function effectively in place of a human-readable “chain of thought.”

纽约大学研究人员在 2024 年发表的一篇论文表明，“无意义的填充标记”（字面意思是一串点）可以有效地代替人类可读的“思想链”。

William Merrill , one of the authors on that paper and currently a professor at the Toyota Technological Institute at Chicago, put the matter plainly: “There’s no guarantee the chain of thought has to be meaningful in any sense.” Pavel Izmailov , a researcher at NYU who also works for Anthropic (and was part of its original reasoning-model team), said he doubts that reinforcement learning — a typical training method for LRMs — even incentivizes models to produce faithful chains of thought in the first place.

该论文的作者之一、现任芝加哥丰田技术学院教授威廉·梅里尔 (William Merrill) 明确指出了这一问题：“无法保证思想链在任何意义上都有意义。”纽约大学研究员 Pavel Izmailov 也为 Anthropic 工作（并且是其最初的推理模型团队的成员），他表示他怀疑强化学习（LRM 的一种典型训练方法）是否会激励模型首先产生忠实的思想链。

“I mean, maybe it will,” he told me.

“我的意思是，也许会，”他告诉我。

“But I would say the chances are not very high.”

“但我想说，可能性不是很大。”

OK, so the linguistic content of reasoning traces may be dubious.

好的，所以推理痕迹的语言内容可能是可疑的。

But surely the tokens themselves must play a role in producing the model’s outputs?

但代币本身肯定在模型输出的生成中发挥作用吗？

(Think of a pinball machine: It runs on coins, not the words “In God We Trust.”)

（想想弹球机：它靠硬币运行，而不是“我们相信上帝”这句话。）

Not so fast.

没那么快。

A 2025 paper from Northeastern University and the University of California, Berkeley on frontier open-source LRMs showed that between 30% and 60% of their “thinking steps” had “minimal causal impact” on the answers the models produced to benchmark math questions.

东北大学和加州大学伯克利分校在 2025 年发表的一篇关于前沿开源 LRM 的论文表明，他们 30% 到 60% 的“思维步骤”对模型为基准数学问题生成的答案具有“最小因果影响”。

Chop half of them out, and a model’s performance barely suffers.

砍掉其中一半，模型的性能几乎不会受到影响。

“We want to be careful when we review these chain-of-thought prompts because they may not be linked to the final output,” said Weiyan Shi , one of the study’s authors.

该研究的作者之一施伟言表示：“我们在审查这些思路提示时要小心，因为它们可能与最终输出无关。”

So reasoning traces, the very things that supposedly distinguish LRMs from the mere next-word-predicting LLMs, are not necessarily either meaningful or causal to a model’s …

因此，推理痕迹（据说将 LRM 与纯粹的下一个词预测 LLM 区分开来的东西）不一定对模型的有意义或因果关系......

reasoning?

推理？

I’m no philosopher, but this seems to stretch the meaning of “reasoning” beyond its tensile strength.

我不是哲学家，但这似乎将“推理”的含义延伸到了其拉伸强度之外。

Kambhampati’s research group sounded frankly fed up in the title of their position paper on the subject (presented at the 2026 International Conference on Machine Learning, one of the field’s most prestigious academic gatherings): “Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces!”

Kambhampati 的研究小组在关于该主题的立场文件（在 2026 年国际机器学习会议上发表，这是该领域最负盛名的学术聚会之一）的标题中坦率地表达了厌倦：“停止将中间标记拟人化为推理/思维痕迹！”

To be clear, Kambhampati, a former president of the Association for the Advancement of Artificial Intelligence, with a background in AI planning algorithms, doesn’t deny that LRMs can work (when they work).

需要明确的是，具有人工智能规划算法背景的人工智能促进协会前主席 Kambhampati 并不否认 LRM 可以发挥作用（当它们发挥作用时）。

“We are in wondrous times,” he told me, when I asked what he thought of OpenAI’s 2026 victory in solving the famous unit distance problem in math .

当我问他对 OpenAI 2026 年解决数学中著名的单位距离问题的胜利有何看法时，他告诉我：“我们正处在一个奇妙的时代。”

If he has a bone to pick, it’s with what he sees as a rush in both academia and industry to embrace overly convenient explanations.

如果说他有什么挑剔的话，那就是他认为学术界和工业界都急于接受过于方便的解释。

A fake theory is worse than admitting that we don’t have a theory.

虚假的理论比承认我们没有理论更糟糕。

Subbarao Kambhampati, Arizona State University

Subbarao Kambhampati，亚利桑那州立大学

“Many ideas that have been proposed [about] the sources of strength [of these models] have been misunderstood or mischaracterized,” he said.

“关于[这些模型]的力量来源提出的许多想法都被误解或错误描述了，”他说。

“There’s this general mindset that says, ‘Let’s go ahead and claim certain abilities, because eventually that might become true anyway.’ And my sense is: That’s not science.

“有一种普遍的心态说，‘让我们继续前进并声称拥有某些能力，因为无论如何最终这可能会成为现实。’我的感觉是：这不是科学。”

That is investment.”

这就是投资。”

On the other side of the AI-reasoning fence, the disdain seems to be mutual.

在人工智能推理围栏的另一边，蔑视似乎是相互的。

“These ‘scientific’ papers from last summer — I would put this in big, big air quotes,” said Sébastien Bubeck , a member of OpenAI’s technical staff (and a prominent evangelist for the company’s reasoning models among scientists and mathematicians).

OpenAI 技术人员（也是该公司推理模型在科学家和数学家中的著名传播者）塞巴斯蒂安·布贝克 (Sébastien Bubeck) 说：“去年夏天的这些‘科学’论文，我会用很大很大的引号来引用它们。”

He called earlier Apple results critiquing AI reasoning “wrong,” claiming that they were due to a training quirk in models that are now obsolete.

他称苹果公司早期的结果批评人工智能推理是“错误的”，并声称这是由于现在已经过时的模型中的训练怪癖造成的。

“Modern models starting with GPT-5.5 do not suffer from this issue,” he said.

“从 GPT-5.5 开始的现代型号不会遇到这个问题，”他说。

“It would be interesting to revisit those results.” (Apple did not make its researchers available for interviews.)

“重新审视这些结果会很有趣。” （苹果公司没有让其研究人员接受采访。）

Here’s the thing: Nobody denies that AI reasoning models can, indeed, produce significant and accurate results.

事情是这样的：没有人否认人工智能推理模型确实可以产生重要且准确的结果。

Furthermore, every researcher I spoke to acknowledged that negative findings about the models’ capabilities on certain reasoning tasks (especially those of smaller, open-source LRMs) may not always generalize to the latest-and-greatest AI products.

此外，我采访过的每一位研究人员都承认，关于模型在某些推理任务（尤其是小型开源 LRM 的模型）上的能力的负面结果可能并不总是适用于最新、最出色的人工智能产品。

Their inner workings remain trade secrets.

他们的内部运作仍然是商业秘密。

But if we’re disinclined (as I am) to simply dismiss contradictory evidence about the mechanisms driving AI reasoning, the question remains: How do we account for it?

但如果我们不愿意（像我一样）简单地驳回关于驱动人工智能推理的机制的矛盾证据，那么问题仍然是：我们如何解释它？

Kambhampati, as it turns out, is interested in doing exactly that.

事实证明，坎巴姆帕蒂对此很感兴趣。

“I’m not negative.

“我并不消极。

I just sound negative because everybody else is way too positive,” he said.

我只是听起来很消极，因为其他人都太积极了，”他说。

“In science, you have to actually understand what the current thing does and what it cannot do.”

“在科学中，你必须真正了解当前事物能做什么和不能做什么。”

One straightforward reason state-of-the-art LRMs work, he told me (a point also echoed by Mitchell), is that they’re often surrounded by “normal” software that guides and verifies their outputs.

他告诉我，最先进的 LRM 工作的一个直接原因（米切尔也赞同这一点）是，它们通常被指导和验证其输出的“正常”软件包围。

Agentic AI systems, which have transformed software engineering since the fall of 2025, work this way.

自 2025 年秋季以来，代理人工智能系统已经改变了软件工程，其工作方式就是这样的。

So does Google DeepMind’s AlphaProof Nexus , which relies on Lean, an automated theorem-proving tool.

Google DeepMind 的 AlphaProof Nexus 也是如此，它依赖于 Lean（一种自动化定理证明工具）。

But Kambhampati is more interested in making sense of stand-alone reasoning models that rely solely on their self-generated reasoning traces — “the ‘think’ part,” he said.

但 Kambhampati 更感兴趣的是理解独立推理模型，这些模型完全依赖于自身生成的推理轨迹——“‘思考’部分，”他说。

The “think” part is what OpenAI, for one, is doubling down on.

其中之一就是 OpenAI 正在加倍努力的“思考”部分。

When I asked Bubeck if the splashy unit distance proof was produced with methods outside the LRM’s own chain of thought — perhaps with Lean verifying its results — he seemed to find the question almost nonsensical.

当我问 Bubeck 引人注目的单位距离证明是否是用 LRM 自己的思想链之外的方法产生的（也许是 Lean 验证了其结果）时，他似乎发现这个问题几乎毫无意义。

“It’s not like we’re making a mystery of it,” he said.

“这并不是说我们要对此神秘化，”他说。

“We have released the chain of thought.

“我们已经释放了思想链。

You can just go and look at it.

你可以去看看。

The whole point is that the model is reasoning like a human would.

重点是该模型像人类一样进行推理。

And when humans reason, we don’t use Lean.” Technically, OpenAI released a “ rewritten summary ” of the model’s chain of thought produced by two human experts using Codex, another OpenAI model.

当人类进行推理时，我们不会使用精益。”从技术上讲，OpenAI 发布了由两位人类专家使用另一个 OpenAI 模型 Codex 制作的模型思想链的“重写摘要”。

Since 2024, the company has not publicly revealed “raw” chains of thought from its reasoning models, a policy also adopted by Google DeepMind and Anthropic.

自 2024 年以来，该公司一直没有公开披露其推理模型中的“原始”思想链，谷歌 DeepMind 和 Anthropic 也采用了这一政策。

Kambhampati’s analysis begins in a surprisingly similar place: with the idea that LRMs are just LLMs with more specific training.

Kambhampati 的分析始于一个惊人的相似之处：LRM 只不过是接受过更具体培训的法学硕士。

“There is no extra magic,” he said.

“没有额外的魔法，”他说。

But he diverges sharply from there.

但他与此截然不同。

“It doesn’t make sense to me that an LLM would actually do a step-by-step description of what it is [reasoning] before giving the solution — because that’s a much harder task than just guessing the solution, given the way that LLMs are trained.”

“对我来说，法学硕士在给出解决方案之前实际上会一步一步地描述它是什么[推理]，这对我来说是没有意义的——因为考虑到法学硕士的训练方式，这比仅仅猜测解决方案要困难得多。”

His working hypothesis is that an LRM, like its LLM precursors, performs what he calls “approximate retrieval” across its vast training corpus: “somewhere in the middle” between pattern matching and reasoning, he said, but closer to the former.

他的工作假设是，LRM 就像它的 LLM 前身一样，在其庞大的训练语料库中执行他所谓的“近似检索”：他说，介于模式匹配和推理之间的“中间位置”，但更接近前者。

The role of “thinking tokens,” then, isn’t to narrate an actual chain of thought (because there isn’t one).

那么，“思考标记”的作用并不是叙述一个实际的思想链（因为不存在这样的思想链）。

Instead, it’s to load up the model’s context window in a way that makes it more likely to predict, or “approximately retrieve,” reasoning-shaped strings of text.

相反，它会加载模型的上下文窗口，使其更有可能预测或“近似检索”推理形状的文本字符串。

Kambhampati compared this process to mumbling words to yourself to jog your memory: It barely matters what the words are (though related ones may help), as long as they knock loose something useful.

坎巴汉帕蒂将这个过程比作对自己默念单词以唤起记忆：单词是什么并不重要（尽管相关的单词可能会有所帮助），只要它们能释放出有用的东西即可。

An LRM’s vast “memory” includes all the call-and-response-like examples of written reasoning it was trained on, mulched into numerical “embeddings” that encode their similarities and differences (plus other inscrutable associations) as geometric relationships in a high-dimensional space.

LRM 庞大的“记忆”包括它所训练的所有类似调用和响应的书面推理示例，并融入数字“嵌入”中，将它们的相似点和差异（以及其他难以理解的关联）编码为高维空间中的几何关系。

Probabilistically arriving at an answer within that space may involve intermediate tokens whose embeddings map to coherent-looking “thoughts” in plain English, but not necessarily.

在该空间内概率性地得出答案可能涉及中间标记，其嵌入映射到简单英语中看起来连贯的“想法”，但不一定。

They could be bits of other languages.

它们可能是其他语言的一部分。

They could be fake exclamations like “aha.” Under the right conditions, they could just be dots.

它们可能是“啊哈”之类的假感叹词。在适当的条件下，它们可能只是点。

You want the right answer for the right reason, so you can trust these things.

您希望因正确的原因获得正确的答案，因此您可以相信这些事情。

Melanie Mitchell, Santa Fe Institute

梅兰妮·米切尔，圣达菲研究所

“Whether the [embedding] actually corresponds to a single word or not” — much less a faithful reasoning process — “is beside the point,” Kambhampati said.

“[嵌入]是否实际上对应于一个单词”——更不用说一个忠实的推理过程——“不是重点，”坎巴汉帕蒂说。

This framing could help explain both the odd “BS”-ness of some chains of thought and the fact that they can elicit accurate outputs anyway.

这个框架可以帮助解释一些思想链奇怪的“BS”性质，以及它们无论如何都能得出准确输出的事实。

It would also neatly account for LRMs’ steady improvement in coding and math — what AI researchers call “verifiable domains.” Code runs, or it doesn’t; proofs are either correct or not.

它还可以巧妙地解释 LRM 在编码和数学方面的稳步改进——人工智能研究人员称之为“可验证领域”。代码运行，或者不运行；证明要么正确，要么不正确。

These binary conditions and the written steps associated with them can create convenient training signals for LRMs.

这些二元条件和与之相关的书面步骤可以为 LRM 创建方便的训练信号。

The model doesn’t have to learn or reliably apply a general reasoning process, Kambhampati said; it just has to absorb enough examples of what the steps look like to predictively mimic them on its way to “stitching together” a plausible result that can then be verified.

Kambhampati 说，该模型不必学习或可靠地应用一般推理过程；它只需要吸收足够多的步骤示例，以预测性地模仿它们，从而“缝合在一起”一个可以验证的合理结果。

The limit of a reasoning model’s training and step-following capability, known as the “inference horizon,” Kambhampati added, was what Apple researchers exposed with their “Illusion of Thinking” paper in 2025.

Kambhampati 补充道，推理模型的训练和步骤跟踪能力的限制（称为“推理视野”）是苹果研究人员在 2025 年发表的“思维幻觉”论文中揭露的。

Newer models have appeared to push this horizon further, albeit jaggedly.

较新的模型似乎进一步推动了这一前景，尽管参差不齐。

“Most of the time they probably are not learning the algorithm” associated with a reasoning process, he said.

他说，“大多数时候，他们可能没有学习与推理过程相关的算法”。

It’s much likelier that they are leveraging an ever-enlarging set of examples and clever reward signals.

他们更有可能利用一组不断扩大的示例和巧妙的奖励信号。

Kambhampati hardly considers his case closed, and neither do I.

坎巴汉帕蒂几乎不认为他的案子已经结束，我也不认为。

But it’s a start — and one I find plausible, given that other researchers have also used similar “it’s the training, stupid” approaches to demystify AI behavior.

但这只是一个开始——我认为这是一个合理的开始，因为其他研究人员也使用了类似的“这是训练，愚蠢的”方法来揭开人工智能行为的神秘面纱。

Still, there was an elephant left in the room: How much does it matter whether or not we can accurately observe, characterize, and validate the processes at work inside large reasoning models?

尽管如此，房间里仍然有一头大象：我们是否能够准确地观察、描述和验证大型推理模型中的工作过程有多重要？

The honest answer, according to Mitchell, is that it depends.

米切尔认为，诚实的答案是，这取决于情况。

“Think of AlphaFold,” she said, referring to Google’s AI tool for predicting protein structures .

“想想 AlphaFold，”她说，指的是谷歌用于预测蛋白质结构的人工智能工具。

“It’s doing some kind of incredibly complex statistical associations.

“它正在进行某种极其复杂的统计关联。

We don’t know what they are, but they seem to work.

我们不知道它们是什么，但它们似乎有效。

These things are [already] black boxes, even without a ‘reasoning trace.’” If LRMs can supercharge mathematics research the way AlphaFold did for computational biology, this line of thinking goes, why not embrace them, idiosyncrasies and all, and just verify the results?

这些东西（已经）是黑匣子了，即使没有‘推理痕迹’。”如果 LRM 能够像 AlphaFold 对计算生物学那样促进数学研究，那么为什么不拥抱它们、特性等等，然后验证结果呢？

“My perspective is: We’re trying to be useful.

“我的观点是：我们正在努力发挥作用。

We’re trying to build these models so that they can solve problems that matter, so that we actually accelerate scientific research,” said Bubeck.

我们正在尝试建立这些模型，以便它们能够解决重要的问题，以便我们真正加速科学研究，”布贝克说。

“It’s more interesting and more productive to talk about what they can do, rather than, ‘Oh, but they can only do that because of X [reasons].’”

“谈论他们能做什么更有趣，也更富有成效，而不是‘哦，但他们只能因为 X [原因] 才能做到这一点。’”

But as Mitchell also points out, the possibility that an LRM could be “right for the wrong reasons” has an obvious relevance to the future of doing research.

但正如米切尔也指出的那样，LRM 可能“因错误的原因而正确”的可能性与研究的未来有着明显的相关性。

“You want the right answer for the right reason, so you can trust these things,” she said, and not just in verifiable domains.

她说：“你想要因正确的原因得到正确的答案，这样你就可以相信这些事情，”而且不仅仅是在可验证的领域。

Tal Linzen , a researcher at NYU and Google whose Computation and Psycholinguistics Lab published results similar to Apple’s “Illusion of Thinking” paper, said that “you want an AI system to be able to apply an algorithm reliably, regardless of whether you call [it] reasoning or not.” Treating chains of thought too reverently — even when their results are verifiable — could also prevent scientists from discovering even better ways of biasing LRMs toward accurate outputs.

纽约大学和谷歌的研究员塔尔·林岑（Tal Linzen）的计算和心理语言学实验室发表了与苹果公司的“思维幻觉”论文类似的结果，他说：“你希望人工智能系统能够可靠地应用算法，无论你是否称之为推理。”过于恭敬地对待思想链——即使它们的结果是可验证的——也可能会阻止科学家发现更好的方法来使 LRM 偏向准确的输出。

“We may be leaving some opportunities unexplored,” said Pradeep Dasigi , a researcher who helped train open LRMs at the Allen Institute for Artificial Intelligence.

“我们可能会留下一些未开发的机会，”艾伦人工智能研究所帮助训练开放式 LRM 的研究员普拉迪普·达西吉 (Pradeep Dasigi) 说。

Kambhampati, unsurprisingly, puts it in even starker terms: Taking the meaning of AI reasoning traces seriously, he said, was a scientific “rabbit hole,” akin to believing in geocentrism or the ether .

不出所料，坎巴汉帕蒂用了更严厉的措辞：他说，认真对待人工智能推理痕迹的意义是一个科学的“兔子洞”，类似于相信地心说或以太。

Harsh, perhaps, but he has a point.

也许很严厉，但他说得有道理。

Those incorrect mental models made intuitive sense at the time, just as chains of thought do now.

那些不正确的心理模型在当时具有直观意义，就像现在的思想链一样。

When an LRM produces a correct answer — along with pages of “thoughts” showing how it got the result — intuition tells us that the two must be linked.

当 LRM 产生正确答案时，以及显示其如何得出结果的“想法”页面时，直觉告诉我们，两者必须联系在一起。

It’s hard to imagine that process and outcome may have little to do with each other.

很难想象过程和结果可能没有什么关系。

But in the 1990s (in an episode Mitchell and Izmailov both brought up), it was hard to imagine how brute-force search could beat world champ Garry Kasparov at chess .

但在 20 世纪 90 年代（在米切尔和伊兹梅洛夫都提到的一集中），很难想象暴力搜索如何能够在国际象棋比赛中击败世界冠军加里·卡斯帕罗夫。

And in 2023, it was hard to intuit how a giant pile of matrix multiplications could write in iambic pentameter .

到了 2023 年，很难凭直觉理解一大堆矩阵乘法如何用抑扬格​​五音步书写。

For most of us, these just weren’t thinkable thoughts.

对于我们大多数人来说，这些都不是可以想象的想法。

Until, suddenly, they were.

直到突然之间，他们就这样了。

In summer 2024, just months before the first LRM appeared, Mitchell turned me on to a concept that I keep returning to in my AI reporting: “wishful mnemonics.” The phrase was first used all the way back in 1976 by the computer scientist Drew McDermott, in a paper with the epically grouchy title “ Artificial Intelligence Meets Natural Stupidity .” I’ll quote the same passage Mitchell did:

2024 年夏天，就在第一个 LRM 出现的几个月前，米切尔让我发现了一个我在人工智能报告中不断提及的概念：“一厢情愿的助记符”。早在 1976 年，计算机科学家德鲁·麦克德莫特 (Drew McDermott) 就在一篇论文中首次使用了这句话，论文的标题极其令人不满：“人工智能遇见自然愚蠢”。我将引用米切尔的同一段话：

A major source of simple-mindedness in AI programs is the use of mnemonics like “UNDERSTAND” or “GOAL” to refer to programs and data structures.

人工智能程序头脑简单的一个主要来源是使用“理解”或“目标”等助记符来指代程序和数据结构。

…

……

If a researcher …

如果研究人员...

calls the main loop of his program “UNDERSTAND,” he is (until proven innocent) merely begging the question.

称他的程序的主循环为“理解”，他（直到被证明是无辜的）只是在回避这个问题。

He may mislead a lot of people, most prominently himself.

他可能会误导很多人，尤其是他自己。

…

……

What he should do instead is refer to this main loop as “G0034,” and see if he can convince himself or anyone else that G0034 implements some part of understanding.

他应该做的是将这个主循环称为“G0034”，看看他是否能让自己或其他人相信 G0034 实现了某些部分的理解。

…

……

Many instructive examples of wishful mnemonics by AI researchers come to mind once you see the point.

一旦你明白了这一点，人工智能研究人员的一厢情愿的助记符的许多有启发性的例子就会浮现在你的脑海中。

This is how I make sense of AI reasoning.

这就是我理解人工智能推理的方式。

LRMs, chains of thought, thinking tokens: It’s wishful mnemonics all the way down — a heady mix of shorthand and suspended disbelief, like Oprah-style “manifesting” with a computer science spin.

LRM、思想链、思维标记：从头到尾都是一厢情愿的助记符——速记和悬而未决的怀疑的令人兴奋的混合，就像奥普拉式的计算机科学旋转的“表现”。

This isn’t necessarily a dig; all novel research likely requires some version of this mindset just to get off the ground.

这不一定是挖掘；所有新颖的研究都可能需要某种形式的这种心态才能起步。

It certainly doesn’t mean AI reasoning can’t or doesn’t work.

这当然并不意味着人工智能推理不能或不起作用。

But the “wishful” part seems to be as powerful as ever.

但“如意”部分似乎一如既往地强大。

“We react to language in a way that is very anthropomorphizing.

“我们对语言的反应非常拟人化。

That’s just the way that we humans work,” Mitchell told me.

这就是我们人类的工作方式，”米切尔告诉我。

Much of the contentious research activity around AI reasoning, she said, “is par for the course.

她说，围绕人工智能推理的许多有争议的研究活动“都是理所当然的。

But in other ways, there’s a lot of very unscientific aspects to it.” Or, as Kambhampati put it, “A fake theory is worse than admitting that we don’t have a theory.”

但在其他方面，它还有很多非常不科学的方面。”或者，正如 Kambhampati 所说，“虚假的理论比承认我们没有理论更糟糕。”

In any case, we have to call it something while we figure out what it is.

无论如何，我们必须在弄清楚它是什么的同时，给它起个名字。

I don’t foresee always reaching for the air quotes around AI reasoning, any more than I’d put them around the “horse” in horsepower.

我并不认为总是会引用有关人工智能推理的空中引言，就像我不会将它们放在马力的“马”周围一样。

LRMs are like engines: They require fuel, emit exhaust, and go fast.

LRM 就像发动机：它们需要燃料、排放废气并且运行速度很快。

Still, when I describe the oomph my Toyota can deliver when I step on the gas, it’s not because I believe there are little hooves pounding away under the hood.

尽管如此，当我描述我的丰田汽车在我踩油门时所能提供的动力时，这并不是因为我相信引擎盖下有小蹄子在猛烈地撞击。

Until a clearer scientific account emerges of what’s going on under the hood of AI reasoning models, I’ll regard their horsepower in a similar spirit — even as the engines roar.

在对人工智能推理模型背后发生的事情有更清晰的科学解释之前，我会以类似的精神看待它们的马力——即使是在引擎轰鸣的时候。

Get highlights of the most important news delivered to your email inbox

获取发送到您的电子邮件收件箱的最重要新闻的摘要

Quanta Magazine moderates comments to facilitate an informed, substantive, civil conversation.

《Quanta》杂志对评论进行审核，以促进知情、实质性、文明的对话。

Abusive, profane, self-promotional, misleading, incoherent or off-topic comments will be rejected.

辱骂、亵渎、自我推销、误导、语无伦次或偏离主题的评论将被拒绝。

Moderators are staffed during regular business hours (New York time) and can only accept comments written in English.

版主在正常工作时间（纽约时间）有工作人员，并且只能接受英文评论。