内容来源：https://www.quantamagazine.org/is-ai-reasoning-right-for-the-wrong-reasons-20260731/

内容总结：

AI“推理”真相扑朔迷离：能力是真的，但过程可能是“胡诌”

近年来，被寄予厚望的AI“推理模型”（LRM）在数学、编程等领域屡创佳绩，甚至能解决悬而未决的数学难题，但与此同时，大量研究却揭示其“推理过程”存在严重问题，这引发了科学界的激烈争论。

一方面，OpenAI、谷歌DeepMind等公司的模型在国际数学奥林匹克竞赛中夺金，甚至在今年5月一举攻克一个著名的开放数学研究问题，展示了惊人的解题能力。但另一方面，来自苹果、圣菲研究所等机构的论文指出，这些模型的“思维链”可能只是“幻觉”，极易在简单条件下“彻底崩溃”，甚至通过“表面捷径”糊弄基准测试。

问题的核心在于：模型输出的那些看似逻辑严谨的“思考步骤”（即思维链），究竟是其真实推理过程的忠实记录，还是一种事后编造的“胡言乱语”？

多所高校的研究发现，这些中间token既非必要，也未必反映模型内部运作。例如，用无关或错误的“思维链”替换掉正确的，模型的最终答案竟不受影响；甚至用一串无意义的省略号“……”也能起到同等效果。亚利桑那州立大学的研究者苏布拉奥·坎巴姆帕蒂直言，这些文本更像是“喃喃自语”，并非真正的推理过程。他的团队在论文标题中直接呼吁“停止将中间token拟人化为推理/思考痕迹”。

对于这些“揭黑”研究，业界也有不同声音。OpenAI技术团队成员塞巴斯蒂安·布贝克反驳称，早前苹果的批评基于已过时的模型，并强调新模型如GPT-5.5已无此问题。他认为，探讨模型“能做什么”远比质疑其“为何能”更有建设性，并指出公司已经发布了模型思维链的“重写摘要”供外界查验。

面对矛盾重重的研究证据，圣菲研究所的梅兰妮·米切尔总结出三点共识：一，推理模型确实提升了任务准确率；二，模型生成的文本并非其内部运作的真实映射；三，这些文本中很大一部分甚至可有可无。

坎巴姆帕蒂提出一种解释：大语言模型的“推理”实质上是基于海量训练数据进行的“近似检索”，介于模式匹配与真正推理之间，且更接近前者。所谓“思考token”的作用，并非叙述一个真实存在的推理链条，而是通过填充上下文窗口，提高模型预测出“像样的推理文本”的概率。这就像自言自语来唤起记忆——具体说什么词不重要，重要的是能“敲打出有用的东西”。

这种“为错而对的推理”对人类科研的启示何在？米切尔指出，无论是蛋白质结构预测的AlphaFold还是数学推理模型，本质上都是高效的黑箱。在可验证的领域（如数学证明、代码运行），AI的“错误推理”或许无伤大雅，因为结果可以校验。但在更广阔的、结果无法自动验证的领域，能否信任一个“动机不纯”的AI，便成了关键问题。纽约大学的塔尔·林岑强调，希望AI系统能可靠地应用算法，而不仅仅是在可验证领域碰运气。

学术界还提醒，过分迷信思维链的“真实性”可能阻碍我们发现更优的模型训练或推理方法。艾伦研究所的普拉迪普·达西吉便担忧，我们可能因此错失许多未探索的机会。

从更宏观的视角看，这种对AI推理的复杂情绪并非新鲜事。计算机科学家德鲁·麦克德莫特早在1976年就批评过AI研究中的“一厢情愿的记忆术”——用“理解”“目标”这类好听的名字去称呼程序内部模块，导致研究者自我误导。如今，大型推理模型、思维链这些概念，在许多人看来正是这种“一厢情愿”的现代版。

尽管如此，没有人否认这些“引擎”能跑出惊人的速度。在当前科学解释尚不清晰的阶段，我们或许只能像看待“马力”中的“马”一样看待AI的“推理”一词——它形象、好用，但若真要刨根问底引擎盖下的机械原理，恐怕还得再等上一等。

中文翻译：

AI“推理”是否用错了理由？

引言

我直说了吧：AI“推理”到底是怎么回事？

抱歉用了引号。这种带引号的侧目态度在2024年还比较常见，当时被称为“大型推理模型”（LRM）的大语言模型近亲才刚刚问世。但如今再这样似乎显得有点无礼了——毕竟2026年5月，OpenAI的一款“通用推理模型”一次性解决了一个著名的开放数学研究问题。尽管如此，面对这些AI系统实际在做什么的科学解释，我依然感到一种智力上的晕头转向，不知该如何表达。

推理有各种技术定义形式，但基本过程很容易辨认：通过把逻辑上环环相扣的中间步骤串联起来，得出一个可靠的结论。我们靠思维完成这一过程；LRM则依靠所谓的“思维链”——这是形容模型在回答复杂问题之前输出的一串合成文本的行话。前一分钟，AI能通过这些思维链进行推理的观点还遭到苹果公司一个研究团队的引人注目且可信的批评，称其为“思维的幻觉”，在出奇简单的条件下就会“彻底崩溃”。下一分钟，LRM就在国际数学奥林匹克竞赛中摘得金牌——这一成就难度极高，“即使是功成名就的数学家和科学家也可能终生将其写进简历”，科学家兼AI批评者加里·马库斯和欧内斯特·戴维斯在2025年如此写道。如果这都不算“真正的”推理，那什么才算？

但等等——不久之后，圣塔菲研究所的更多研究表明，LRM即使面对精心设计的推理基准测试（比如一组类类比视觉谜题），也能仅仅靠“表面层面的‘捷径’”轻松碾压。它们的行为看起来与其说是可泛化的推理，不如说是在钻系统的空子。然后，仿佛掐准了时机，又一个“接下来看我的”时刻降临：谷歌DeepMind和数学家陶哲轩（史上最伟大！）利用AI重新发现或改进了67个问题的解法，涵盖“数学分析、组合学、几何和数论”。不服来辩，黑子们！

那又怎么解释其他证据表明LRM即使拥有必要的算法和计算预算，也无法可靠地推理，并且存在一堆足以铺成一条嬉水滑道的、经科学记录的各种失败模式呢？随它去吧——我猜这就是所谓的“锯齿形智能”吧（AI行话，意思是“能用的时候就好用”）。

从2025年末到2026年，事态就这样反复拉锯。我做科学记者已有20年，其中一半时间在报道AI，所以我知道不该指望日新月异的研究能保持整齐划一的一致性。但即使对我而言，这种来回折腾也有点过火了。借用阿尔·帕西诺在《惊爆内幕》中的台词：“我现在有两种感受：愤怒，和好奇。”我不认为这里存在欺诈。我只想知道方向到底在哪边。AI推理能不能同时既是“扯淡”又不是“扯淡”？如果能，那到底是怎么运作的？

我知道该先打给谁。

梅兰妮·米切尔在AI领域的职业生涯可以追溯到20世纪80年代，但近年来她以紧跟前沿的AI真相讲述者而闻名，为《科学》杂志和她订阅量颇高的通讯撰写通俗易懂的解说文章，同时在圣塔菲研究所从事研究。（那篇关于“表面层面的‘捷径’”的研究就是她的。）当我问她关于AI推理我们究竟知道些什么时，她的回答短得足以写在一张索引卡上。

“第一：它有效。它改善了结果，”她说，指的是LRM在推理任务上的准确率优于LLM。“第二：实际生成的文本”——也就是每个LRM都被训练出来以提高性能的思维链——“不一定是模型内部实际发生情况的忠实反映。第三：这些文本中有很多甚至没用。你其实可以把它去掉。”

我们来拆解第二点和第三点，因为“既扯淡又不扯淡”的叠加态就住在这里。思维链在2022年被半发现、半设计出来，最初是LLM的一种提示技巧：给它们提供书面推理的示例（或者，著名的方法是，只需让它们“一步一步地思考”），它们就会在简单的逻辑和数学问题上给出不那么愚蠢的答案。从2024年OpenAI的o1模型开始，LRM通过生成这类提示——也被称为推理轨迹或思考令牌——然后将其反馈给自己，来训练实现这一技巧的自动化。因为LRM本质上只是语言模型，这些额外的文本片段制造出了一种看起来很像模型“思维过程”的书面记录。

但事情没这么简单。越来越多的学术和行业研究表明，这些“中间令牌”是否忠实地代表了LRM的内部运作，是值得怀疑的。它们不是可审计的收据或准确的报告，而更像是亚利桑那州立大学研究员苏巴拉奥·坎巴姆帕蒂所说的“喃喃自语”——确实是语言片段，但其含义可能与发生的任何推理完全无关。坎巴姆帕蒂的实验室在2025年证明，将模型正确的“轨迹”完全替换为错误或不相关的轨迹，并不会降低它在形式推理任务上的表现。与此同时，仅用正确的轨迹数据训练模型，仍然会导致它偶尔生成无效的推理记录——即使它对给定的原始问题产生了正确的解答。2024年纽约大学研究人员发表的一篇论文表明，“无意义的填充令牌”——字面意义上的点点点字符串——可以有效地替代人类可读的“思维链”。

威廉·梅里尔是那篇论文的作者之一，目前在芝加哥丰田技术研究所任教授，他把话说得很直白：“没有任何保证说思维链必须在任何意义上是有意义的。”NYU研究员帕维尔·伊兹梅洛夫同时在Anthropic工作（并且是其原始推理模型团队的一员），他说他怀疑强化学习——LRM的一种典型训练方法——甚至从一开始就不会激励模型产生忠实的思维链。“我的意思是，也许它会，”他告诉我，“但我得说几率不是很高。”

好吧，所以推理轨迹的语言内容可能是可疑的。但那些令牌本身肯定在产生模型输出的过程中发挥了作用吧？（想想弹球机：它靠硬币运行，而不是靠“我们信上帝”这几个字。）

别急。东北大学和加州大学伯克利分校2025年的一篇关于前沿开源LRM的论文表明，它们30%到60%的“思考步骤”对模型在基准数学题上产生的答案“因果影响极小”。砍掉一半，模型的表现几乎不受影响。“我们在审阅这些思维链提示时要谨慎，因为它们可能与最终输出没有关联，”该研究的作者之一石薇言说。

那么推理轨迹——据称正是它们将LRM区别于单纯的下一词预测型LLM——既不必然有意义，也不必然是模型……推理的因果因素？我不是哲学家，但这似乎把“推理”的含义拉伸到了超过其抗拉强度的程度。坎巴姆帕蒂的研究团队在他们关于这一主题的立场文件标题中简直显得受够了（该文件发表在2026年国际机器学习会议——该领域最负盛名的学术会议之一——上）：“停止将中间令牌拟人化为推理/思考轨迹！”

需要说明的是，坎巴姆帕蒂是人工智能促进协会的前主席，背景是AI规划算法，他并不否认LRM能工作（在它们能工作的时候）。“我们身处奇妙的时代，”当我问他如何看待OpenAI在2026年解决数学中著名的单位距离问题的胜利时，他这样告诉我。如果说他有什么要吐槽的，那就是他认为学术界和工业界都在急于拥抱过于方便的解读。

“关于这些模型优势来源的许多提议都被误解或曲解了，”他说，“有一种普遍的心态说：‘咱们就先宣称某些能力吧，因为反正最终可能会成真。’而我的感觉是：那不是科学。那是投资。”

在AI推理栅栏的另一边，不屑似乎是相互的。“去年夏天那些‘科学’论文——我要给‘科学’加上大大的引号，”OpenAI技术团队成员塞巴斯蒂安·布贝克说（他也是该公司推理模型在科学家和数学家中的重要布道者）。他称苹果此前批评AI推理的结果是“错误的”，声称那些问题源于已过时模型的一个训练怪癖。“从GPT-5.5开始的新模型不存在这个问题，”他说，“重新审视那些结果会很有趣。”（苹果未让其研究人员接受采访。）

事情是这样的：没有人否认AI推理模型确实能产生重大且准确的结果。此外，我所交谈的每一位研究人员都承认，关于模型在特定推理任务上能力的负面发现（尤其是较小规模的开源LRM），可能并不总是适用于最新最强大的AI产品。它们的内在工作机制仍是商业机密。但如果我们不愿意（像我一样）简单地将关于AI推理机制相互矛盾的证据一丢了之，那么问题依然存在：我们该如何解释它？

事实证明，坎巴姆帕蒂正有兴趣做这件事。“我并不悲观。我只是听起来悲观，因为其他人都太乐观了，”他说，“在科学中，你必须真正理解当前的东西能做什么、不能做什么。”

他告诉我，最先进的LRM有效的一个直接原因（米切尔也赞同这一点）是，它们通常被“常规”软件包围，这些软件指导和验证它们的输出。智能体AI系统——自2025年秋以来彻底改变了软件工程——就是这样运作的。谷歌DeepMind的AlphaProof Nexus也是如此，它依赖Lean这个自动化定理证明工具。但坎巴姆帕蒂更感兴趣的是理解那些完全依赖自生成推理轨迹的独立推理模型——“即‘思考’的部分，”他说。

“思考”部分正是OpenAI等公司正在加码的。当我问布贝克，那个引人注目的单位距离证明是否使用了LRM自身思维链之外的方法——比如Lean来验证结果——他似乎觉得这个问题几乎毫无意义。

“我们并没有把它搞得神秘兮兮，”他说，“我们已经发布了思维链。你直接去看就行。重点在于模型像人类一样推理。而人类推理时，我们不会用Lean。”严格来说，OpenAI发布的是由两位人类专家使用Codex（另一款OpenAI模型）生成的模型思维链的“改写摘要”。自2024年以来，该公司就没有公开披露其推理模型的“原始”思维链，谷歌DeepMind和Anthropic也采取了同样的政策。

坎巴姆帕蒂的分析从一个惊人相似的地方开始：LRM只是经过更专门训练的LLM。“没有额外的魔法，”他说。但他在此之后便急剧分道扬镳。“在我看来，一个LLM真的会在给出解答之前逐步描述它在推理什么，这说不通——因为按照LLM的训练方式，这比直接猜测答案要难得多。”

他的工作假设是，LRM和它的LLM前辈一样，在其庞大的训练语料上进行他所称的“近似检索”：介于模式匹配和推理之间的某个位置，但更接近前者。那么“思考令牌”的角色就不是叙述一条真实的思维链（因为根本不存在），而是以某种方式加载模型的上下文窗口，使其更有可能预测或“近似检索”出具有推理形态的文本串。

坎巴姆帕蒂把这个过程比作自己喃喃自语来唤起记忆：说的词是什么几乎无所谓（尽管相关的词可能有帮助），只要它们能激发出有用的东西就行。LRM庞大的“记忆”包括它训练时见过的所有问答式书面推理示例，被搅碎成数值“嵌入”，把它们的相似性和差异性（以及其他难以捉摸的关联）编码为高维空间中的几何关系。在这个空间中概率性地得出答案，可能涉及那些嵌入映射为看起来连贯的英文“想法”的中间令牌——但不一定。它们可能是其他语言的片段。它们可能是“啊哈”这样的假感叹。在适当条件下，它们甚至可能只是点。

“（嵌入）是否真的对应一个词”——更不用说一个忠实的推理过程了——“并不是重点，”坎巴姆帕蒂说。

这个框架有助于解释某些思维链奇怪的“扯淡”性质，以及它们仍然能引出准确输出的事实。它也能很好地解释LRM在编程和数学上的稳步提升——即AI研究者所说的“可验证领域”。代码能跑就是能跑，不能就是不能；证明要么正确要么错误。这些二元条件和与之相关的书面步骤可以为LRM提供便捷的训练信号。坎巴姆帕蒂说，模型不必学习或可靠地应用一种通用推理过程；它只需要吸收足够多的步骤示例，以便在走向“缝合”一个可验证的合理结果的过程中预测性地模仿它们。

坎巴姆帕蒂补充说，推理模型训练和步骤遵循能力的极限——即所谓的“推理地平线”——正是苹果研究人员在2025年的“思维的幻觉”论文中暴露出来的。新模型似乎在推动这一地平线更远，尽管是锯齿状地推进。“大多数时候它们可能并没有学到与推理过程相关的算法，”他说，更有可能的是它们在利用不断扩大的示例集和巧妙设计的奖励信号。

坎巴姆帕蒂并不认为他的论证已经盖棺定论，我也不这么认为。但这至少是一个起点——而且我觉得它可信，因为其他研究者也曾用类似的“关键在于训练，笨蛋”的思路来解释AI行为。不过，房间里还有一头大象：我们能否准确观察、描述和验证大型推理模型内部运作过程，这件事到底有多重要？

据米切尔说，诚实的答案是：看情况。“想想AlphaFold，”她说，指的是谷歌预测蛋白质结构的AI工具，“它在做某种极其复杂的统计关联。我们不知道那些关联是什么，但它们似乎有效。这些东西本来就是黑箱，即使没有‘推理轨迹’也一样。”如果LRM能像AlphaFold对计算生物学那样为数学研究注入超强动力，这种思路认为，为什么不拥抱它们，连同它们的怪癖一起，只需验证结果就行？“我的观点是：我们努力做到有用。我们构建这些模型是为了让它们解决重要的问题，从而真正加速科学研究，”布贝克说，“讨论它们能做什么更有趣、更有成效，而不是‘哦，但它们能做到只是因为X（原因）。’”

但正如米切尔也指出的，LRM可能“因为错误的原因而得出正确结果”这一可能性，对科研的未来有着显而易见的重要意义。“你想要的是出于正确理由的正确答案，这样才能信任这些东西，”她说，而不仅仅是在可验证的领域。

塔尔·林岑是NYU和谷歌的研究员，他的计算与心理语言学实验室发表了与苹果“思维的幻觉”论文类似的结果。他说：“无论你称之为推理还是不推理，你都希望AI系统能够可靠地应用一种算法。”对思维链过于虔诚——即使结果可以验证——也可能阻止科学家发现更好的方式，使LRM更偏向准确输出。“我们可能在留下一些机会未被探索，”艾伦人工智能研究所研究员普拉迪普·达西吉说，他曾帮助训练开源LRM。不出所料，坎巴姆帕蒂用更尖锐的话说：认真对待AI推理轨迹的含义，他说，是一条科学的“兔子洞”，就像相信地心说或以太一样。

也许严厉了些，但他说得有道理。那些错误的心智模型在当时显得合乎直觉，就像现在思维链看起来合乎直觉一样。当一个LRM产生正确答案——连同数页展示如何得出结果的“思考”——直觉告诉我们两者必有联系。很难想象过程和结果可能几乎互不相干。但在20世纪90年代（米切尔和伊兹梅洛夫都提到过这个插曲），很难想象暴力搜索能在国际象棋中击败世界冠军加里·卡斯帕罗夫。而在2023年，很难凭直觉想象一大堆矩阵乘法能用抑扬格五音步写诗。对我们大多数人来说，这些在当时都是不可想象的想法。直到有一天，它们突然就变成现实了。

2024年夏天，就在第一款LRM问世前几个月，米切尔向我介绍了一个概念，我在AI报道中反复回到它：“一厢情愿的助记符”。这个短语最早出现在1976年，由计算机科学家德鲁·麦克德莫特在一篇题目极其暴躁的论文《人工智能遇上天然的愚蠢》中使用。我引用米切尔引用过的同一段话：

AI程序中思维简单化的一个主要来源是使用“理解”或“目标”这样的助记符来指代程序和数据结构。……如果一位研究者……把他程序的主循环叫作“理解”，他（在证明清白之前）只是在回避问题。他可能会误导很多人，最重要的是误导他自己。……他应该做的是把这个主循环叫作“G0034”，然后看看能否说服自己或别人G0034实现了理解的某个部分。……一旦你明白了这一点，AI研究者们许多关于一厢情愿的助记符的启发性例子就纷至沓来。

这就是我理解AI推理的方式。LRM、思维链、思考令牌：一路下去全是“一厢情愿的助记符”——一种简称和悬置怀疑的刺激混合体，就像奥普拉风格的“显化”加上计算机科学的调料。这不一定是在挖苦；所有新颖的研究可能都需要某种版本的心态才能起步。这当然不意味着AI推理不能或不会发挥作用。但“一厢情愿”的那部分似乎和以往一样强大。

“我们对语言的反应方式非常拟人化。这就是我们人类运作的方式，”米切尔告诉我。围绕AI推理的大量争议性研究活动，她说，“大多在意料之中。但在其他方面，它有很多非常不科学的地方。”或者，正如坎巴姆帕蒂所说：“一个虚假的理论比承认我们没有理论更糟。”

无论如何，在搞清楚它是什么之前，我们总得给它起个名字。我不认为自己会永远给AI推理加上引号，就像我不会给“马力”中的“马”加引号一样。LRM就像发动机：需要燃料，排放废气，跑得飞快。尽管如此，当我描述我的丰田踩下油门时能输出的动力时，并不是因为我相信引擎盖下有小小的马蹄在奔跑。在关于AI推理模型引擎盖下到底发生了什么出现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——即使引擎在轰鸣。

英文来源：

Is AI Reasoning Right for the Wrong Reasons?
Introduction
I’ll just say it: What the hell is going on with AI “reasoning”?
Sorry for the air quotes. 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. 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. Still, I’m not sure how else to acknowledge my intellectual whiplash over the scientific interpretation of what these AI systems are actually doing.
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. 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. 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. 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. If that’s not a sign of “real” reasoning, what is?
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. Then, as if on cue, another “hold my beer” moment: Google DeepMind and the mathematician Terence Tao (the GOAT!) used AI to rediscover or improve the solutions to 67 problems “spanning mathematical analysis, combinatorics, geometry, and number theory.” Deal with it, haters!
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? Whatever — I guess that’s just “jagged intelligence” for you (AI-speak for “when it works, it works”).
And so it went from late 2025 into 2026. 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. But even for me, this back-and-forth has been a bit much. 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. I just want to know which way is up. Can AI reasoning somehow be both BS and not at the same time? And if so, how on Earth does that work?
I knew just who to call first.
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. (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.
“Number one: It works. It improves things,” she said, referring to LRMs’ superior accuracy on reasoning tasks compared to LLMs. “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]. And number three: A lot of that text isn’t even useful. You can actually take it out.”
Let’s unpack numbers two and three, because that’s where the superposition of “BS and not” actually lives. 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. 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. 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.”
Except it’s not that simple. 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. 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. 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. 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. 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.”
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. “I mean, maybe it will,” he told me. “But I would say the chances are not very high.”
OK, so the linguistic content of reasoning traces may be dubious. But surely the tokens themselves must play a role in producing the model’s outputs? (Think of a pinball machine: It runs on coins, not the words “In God We Trust.”)
Not so fast. 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. Chop half of them out, and a model’s performance barely suffers. “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.
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 … reasoning? I’m no philosopher, but this seems to stretch the meaning of “reasoning” beyond its tensile strength. 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!”
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). “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. 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.
“Many ideas that have been proposed [about] the sources of strength [of these models] have been misunderstood or mischaracterized,” he said. “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. That is investment.”
On the other side of the AI-reasoning fence, the disdain seems to be mutual. “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). He called earlier Apple results critiquing AI reasoning “wrong,” claiming that they were due to a training quirk in models that are now obsolete. “Modern models starting with GPT-5.5 do not suffer from this issue,” he said. “It would be interesting to revisit those results.” (Apple did not make its researchers available for interviews.)
Here’s the thing: Nobody denies that AI reasoning models can, indeed, produce significant and accurate results. 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. Their inner workings remain trade secrets. 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?
Kambhampati, as it turns out, is interested in doing exactly that. “I’m not negative. I just sound negative because everybody else is way too positive,” he said. “In science, you have to actually understand what the current thing does and what it cannot do.”
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. Agentic AI systems, which have transformed software engineering since the fall of 2025, work this way. So does Google DeepMind’s AlphaProof Nexus, which relies on Lean, an automated theorem-proving tool. 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.
The “think” part is what OpenAI, for one, is doubling down on. 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.
“It’s not like we’re making a mystery of it,” he said. “We have released the chain of thought. You can just go and look at it. The whole point is that the model is reasoning like a human would. 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. 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.
Kambhampati’s analysis begins in a surprisingly similar place: with the idea that LRMs are just LLMs with more specific training. “There is no extra magic,” he said. But he diverges sharply from there. “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.”
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. The role of “thinking tokens,” then, isn’t to narrate an actual chain of thought (because there isn’t one). 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.
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. 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. 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. They could be bits of other languages. They could be fake exclamations like “aha.” Under the right conditions, they could just be dots.
“Whether the [embedding] actually corresponds to a single word or not” — much less a faithful reasoning process — “is beside the point,” Kambhampati said.
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. 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. These binary conditions and the written steps associated with them can create convenient training signals for LRMs. 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.
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. Newer models have appeared to push this horizon further, albeit jaggedly. “Most of the time they probably are not learning the algorithm” associated with a reasoning process, he said. It’s much likelier that they are leveraging an ever-enlarging set of examples and clever reward signals.
Kambhampati hardly considers his case closed, and neither do I. 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. 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?
The honest answer, according to Mitchell, is that it depends. “Think of AlphaFold,” she said, referring to Google’s AI tool for predicting protein structures. “It’s doing some kind of incredibly complex statistical associations. We don’t know what they are, but they seem to work. 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? “My perspective is: We’re trying to be useful. We’re trying to build these models so that they can solve problems that matter, so that we actually accelerate scientific research,” said Bubeck. “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].’”
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. “You want the right answer for the right reason, so you can trust these things,” she said, and not just in verifiable domains.
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. “We may be leaving some opportunities unexplored,” said Pradeep Dasigi, a researcher who helped train open LRMs at the Allen Institute for Artificial Intelligence. 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.
Harsh, perhaps, but he has a point. Those incorrect mental models made intuitive sense at the time, just as chains of thought do now. 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. It’s hard to imagine that process and outcome may have little to do with each other. 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. And in 2023, it was hard to intuit how a giant pile of matrix multiplications could write in iambic pentameter. For most of us, these just weren’t thinkable thoughts. Until, suddenly, they were.
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:
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. … If a researcher … calls the main loop of his program “UNDERSTAND,” he is (until proven innocent) merely begging the question. He may mislead a lot of people, most prominently himself. … 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. … Many instructive examples of wishful mnemonics by AI researchers come to mind once you see the point.
This is how I make sense of AI reasoning. 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. This isn’t necessarily a dig; all novel research likely requires some version of this mindset just to get off the ground. It certainly doesn’t mean AI reasoning can’t or doesn’t work. But the “wishful” part seems to be as powerful as ever.
“We react to language in a way that is very anthropomorphizing. That’s just the way that we humans work,” Mitchell told me. Much of the contentious research activity around AI reasoning, she said, “is par for the course. 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.”
In any case, we have to call it something while we figure out what it is. 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. LRMs are like engines: They require fuel, emit exhaust, and go fast. 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. 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.