[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-lrm-chain-of-thought-causal-impact-2026":3,"topics-all":35,"news-related-89e317cd-59e8-498c-9be7-bdccb71a69c1":54},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"89e317cd-59e8-498c-9be7-bdccb71a69c1","Quanta 长文复盘 LRM 推理真相：30%–60% 思维链步骤对答案几乎无因果影响","Quanta Magazine 7 月 31 日长文梳理大型推理模型（LRM）思维链是否真反映内部机理：东北大学与 UC Berkeley 联合研究发现 30%–60% 思维链步骤对答案\"几乎没有因果影响\"，Mitchell、Kambhampati 等学者呼吁停止把推理轨迹神化。","# Quanta 长文复盘 LRM 推理真相：30%–60% 思维链步骤对答案几乎无因果影响\n\n当 OpenAI 在 2026 年 5 月用通用推理模型一举解决困扰数学界多年的单位距离猜想时，\"AI 真的会推理\"这件事看上去已经盖棺定论。但 Quanta Magazine 7 月 31 日刊出的长文《Is AI Reasoning Right for the Wrong Reasons?》把这种乐观拆开重审，结论直白：LRM（Large Reasoning Model，大型推理模型）的\"思维链\"很可能只是模型对自己的一种喃喃自语，并不忠实反映内部运算。[原文](https:\u002F\u002Fwww.quantamagazine.org\u002Fis-ai-reasoning-right-for-the-wrong-reasons-20260731\u002F)\n\n## 一、核心数据：30%–60% 的步骤可被裁掉而答案几乎不变\n\n文章援引 2025 年东北大学（Northeastern University）与加州大学伯克利分校（UC Berkeley）联合发表的一篇 [arXiv 论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2510.24941)——他们对开源前沿 LRM 在数学基准题上的\"思考步骤\"做了因果分析，发现 **30%–60% 的思考步骤对最终答案\"几乎没有因果影响\"**。换句话说，把这些步骤随机删掉，模型在数学题上的成绩几乎不动。\n\n该论文作者之一 Weiyan Shi 接受 Quanta 采访时直言：\"我们审阅这些思维链提示时要小心，因为它们可能与最终输出并没有真正的因果联系。\"\n\n## 二、三位关键人物的判断\n\n文章围绕三个学术权威的判断展开：\n\n- **Melanie Mitchell**（圣塔菲研究所）：她的三句话总结最常被引用——\"第一，它有效，确实提升了准确率；第二，模型生成的那串文本不一定忠实反映模型内部在做什么；第三，这些文本很多时候甚至没用，你完全可以把它们删掉。\"\n- **Subbarao Kambhampati**（亚利桑那州立大学，前 AAAI 主席）：他的实验室 2025 年证明，把模型正确的\"推理轨迹\"完全替换成错误或无关文本，模型在形式推理任务上的表现并不会下降；反之，只用正确轨迹训练，模型仍会生成与自身推理不一致的记录。他在 2026 年 ICML 提交了立场论文，标题很直白——\"Stop Anthropomorphizing Intermediate Tokens as Reasoning\u002FThinking Traces!\"（停止把中间 token 拟人化成推理\u002F思考轨迹）。\n- **William Merrill**（芝加哥丰田技术研究所）：作为 2024 年那篇\"用乱码点号代替思维链仍能工作\"的 NYU 论文作者之一，他的判断更狠——\"我们完全没有理由认为思维链必须是有意义的。\"\n\n## 三、机制猜测：思维链不是\"步骤\"而是\"上下文填充\"\n\nKambhampati 给出的工作假说是：LRM 本质上仍是 LLM，只是被训练得更专精。所谓\"思维链\"的作用不是把\"一步推一步\"的过程讲清楚，而是给模型的上下文窗口塞进足够多\"看起来像推理\"的 token，让它更有可能在概率空间中\"近似检索\"出推理形态的答案。把它比作自言自语帮自己回忆——只要把话头抖出来，词具体是什么反而无所谓。Pavel Izmailov（NYU，也在 Anthropic 推理模型团队）补充说，RL 训练并没有真正激励模型产出\"忠实\"的思维链，\"有可能以后会，但我会说概率不高。\"\n\n这个解释能兼容 LRM 在 IMO 金牌、ARC-AGI 等\"可验证领域\"的成功——代码能跑、证明可对，二值信号天然适合给 LRM 当训练奖励。它并不需要真的学会通用推理过程，只要见过足够多\"推理的样子\"并学会模仿。\n\n## 四、反方意见\n\nOpenAI 的 Sébastien Bubeck 反驳说，2025 年 Apple 那篇《Illusion of Thinking》暴露的问题源于已过时的训练技巧，\"从 GPT-5.5 开始的现代模型不会再有这个问题\"。文章同时承认，每位受访学者都指出：关于小尺寸开源 LRM 的负面结论不一定能外推到最新闭源产品，因为这些产品的内部机制仍是商业秘密。\n\n## 五、评论：别再把\"思考\"当成可审计的证据\n\nMitchell 的反问值得整个行业记住：\"你想让 AI 因为正确的理由给出正确的答案，这样你才能信任它。\"这才是 LRM 真正应该被审视的尺度——不是说它\"没用\"，而是说它的\"过程\"目前不是一种可解释、可追责、可审计的证据。\n\n对工程团队来说，这意味着：\n\n- **不要把 LRM 的思维链直接当成审计日志**。它能解释答案的概率并不高。\n- **可验证领域的成绩（数学、代码）不能外推到开放推理**。代码能跑≠一般推理可信。\n- **可解释性研究值得更多投入**。Mitchell 多次强调，对 AlphaFold 那种\"黑盒但有效\"我们可以接受，但如果 LRM 要进入科研、医疗、法律等高风险场景，\"对的理由 + 对的答案\"才是可用的标准。\n\n来源：[Quanta Magazine 长文](https:\u002F\u002Fwww.quantamagazine.org\u002Fis-ai-reasoning-right-for-the-wrong-reasons-20260731\u002F)、[Northeastern & UC Berkeley arXiv 论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2510.24941)、[Kambhampati 立场论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.09762)、[Merrill NYU 论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.15758)。","https:\u002F\u002Fwww.quantamagazine.org\u002Fis-ai-reasoning-right-for-the-wrong-reasons-20260731\u002F","893aafdb-4307-4297-9040-18ca89e0f49a",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"41afa1b4-59d8-49f0-b1d2-4efbccf29b2d","en","Quanta: 30-60% of reasoning steps have no causal effect","Quanta Magazine's July 31 long read revisits whether large reasoning models actually \"reason\" via chain of thought. Citing a 2025 Northeastern \u002F UC Berkeley study, the article reports 30%–60% of chain-of-thought steps have minimal causal impact on the final answer. Melanie Mitchell, Subbarao Kambhampati and others argue the field should stop treating reasoning traces as faithful evidence.","# Quanta Magazine Re-Examines LRM Reasoning: 30%–60% of Chain-of-Thought Steps Have Near-Zero Causal Impact on Answers\n\nWhen OpenAI's general-purpose reasoning model settled the long-standing unit distance problem in May 2026, the question \"Can AI really reason?\" seemed settled. But Quanta Magazine's July 31 feature, *Is AI Reasoning Right for the Wrong Reasons?*, takes that optimism apart. Its conclusion is blunt: a large reasoning model's (LRM's) \"chain of thought\" is most likely the model muttering to itself, not a faithful record of its internal computation. [Original article](https:\u002F\u002Fwww.quantamagazine.org\u002Fis-ai-reasoning-right-for-the-wrong-reasons-20260731\u002F)\n\n## 1. The headline data: 30%–60% of steps can be removed with almost no change\n\nThe piece cites a 2025 [paper from Northeastern University and UC Berkeley](https:\u002F\u002Farxiv.org\u002Fabs\u002F2510.24941) that ran causal analysis on the \"thinking steps\" of frontier open-source LRMs on math benchmarks. Their finding: **30%–60% of the thinking steps had \"minimal causal impact\" on the final answer**. Chop them out, and a model's math score barely budges.\n\nCo-author Weiyan Shi told Quanta: \"We want to be careful when we review these chain-of-thought prompts because they may not be linked to the final output.\"\n\n## 2. What the three leading voices think\n\nThe story is structured around three academic authorities:\n\n- **Melanie Mitchell** (Santa Fe Institute) supplies the most-cited three-sentence summary: \"Number one: it works, it improves things. Number two: the actual text that's generated isn't necessarily faithful to what's going on inside the model. Number three: a lot of that text isn't even useful. You can actually take it out.\"\n- **Subbarao Kambhampati** (Arizona State University, former AAAI president) and his lab showed in 2025 that fully replacing a model's \"correct\" traces with incorrect or irrelevant ones does not degrade performance on a formal reasoning task. Even when trained only on correct traces, a model still occasionally generates invalid records of its reasoning. His 2026 ICML position paper has an unambiguous title: \"Stop Anthropomorphizing Intermediate Tokens as Reasoning\u002FThinking Traces!\"\n- **William Merrill** (Toyota Technological Institute at Chicago), co-author of the 2024 NYU paper showing that \"meaningless filler tokens\" — literal strings of dots — can function in place of a human-readable chain of thought, goes further: \"There's no guarantee the chain of thought has to be meaningful in any sense.\"\n\n## 3. The mechanism: the chain of thought fills the context window, not the steps\n\nKambhampati's working hypothesis is that an LRM is still fundamentally an LLM, just trained to a more specific objective. The \"thinking tokens\" are not there to lay out a true step-by-step proof. Their job is to load up the model's context window with enough \"reasoning-shaped\" text so that it is more likely to probabilistically retrieve an answer that looks like reasoning. It is like muttering to yourself to jog your memory: the specific words barely matter, as long as they shake something useful loose. Pavel Izmailov (NYU, also on Anthropic's original reasoning-model team) adds that RL training may not even incentivize faithful chains: \"I would say the chances are not very high.\"\n\nThis framing is consistent with LRM successes in \"verifiable domains\" such as IMO gold medals and ARC-AGI: code runs or it does not; a proof is correct or not. Those binary signals are convenient training rewards. The model does not have to learn a general reasoning process — it just has to absorb enough examples of what the steps *look like*, then mimic them in order to \"stitch together\" a plausible result that can be checked.\n\n## 4. The counter-argument\n\nOpenAI's Sébastien Bubeck pushes back: the 2025 Apple \"Illusion of Thinking\" findings were an artifact of outdated training tricks. \"Modern models starting with GPT-5.5 do not suffer from this issue.\" The piece also concedes that every interviewed researcher notes the negative findings on smaller open-source LRMs may not generalize to the latest closed-source products, whose internals remain trade secrets.\n\n## 5. Commentary: stop treating \"thinking\" as auditable evidence\n\nMitchell's framing is the one the whole field should sit with: \"You want the right answer for the right reason, so you can trust these things.\" That is the real standard against which LRMs should be judged — not \"does it work,\" but \"is the process it shows me something I can stand behind.\"\n\nFor engineering teams this means:\n\n- **Do not treat the LRM chain of thought as an audit log.** It is unlikely to faithfully explain the answer.\n- **Verifiable-domain wins (math, code) do not generalize to open-ended reasoning.** \"Code runs\" is not the same as \"general reasoning is trustworthy.\"\n- **Interpretability research deserves more investment.** We can accept black boxes à la AlphaFold when the results are independently verifiable, but for LRMs to enter high-stakes domains — scientific research, medicine, law — \"right answer for the right reason\" must become the bar.\n\nSources: [Quanta Magazine feature](https:\u002F\u002Fwww.quantamagazine.org\u002Fis-ai-reasoning-right-for-the-wrong-reasons-20260731\u002F), [Northeastern & UC Berkeley arXiv paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2510.24941), [Kambhampati position paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2504.09762), [NYU filler-tokens paper by Merrill et al.](https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.15758).","lrm-chain-of-thought-causal-impact-2026","2026-08-03T14:50:00Z","2026-08-03T12:50:04.358868Z","2026-08-03T12:50:04.358878Z",true,"agent",229,[36,45],{"slug":37,"tag_slug":37,"title_zh":38,"title_en":39,"intro_zh":40,"intro_en":41,"id":42,"is_active":32,"created_at":43,"modified_at":44},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":46,"tag_slug":46,"title_zh":47,"title_en":48,"intro_zh":49,"intro_en":50,"id":51,"is_active":32,"created_at":52,"modified_at":53},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":55},[56,61,66,71,76,81],{"id":57,"title":58,"news_slug":59,"published_at":60},"0141e95e-e10e-4d88-bf4a-4d1e7df9bfaa","十分钟用 AI 就让人失去攻克难题的意志力","ten-minutes-ai-erodes-persistence-hard-tasks","2026-10-16T03:00:00+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"e12f8e50-3f5f-478c-a099-42518eb2f5c1","用 AI 仅十分钟,专注力和意志力就被削弱","ai-ten-minutes-erodes-persistence","2026-10-13T03:00:00+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"fc533af5-9e8a-43c2-8408-2f931a6aa398","EngramEdit:把事实塞进 LLM 的记忆抽屉","engramedit-decoupled-knowledge-llm","2026-10-09T04:00:00+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"c368ad9f-9308-4a0b-8f5c-3ae4601b48b9","D2K-Bench: 专家设计把 LLM 写 GPU 核提速 33.9%","d2k-bench-llm-gpu-kernel-design-guidance","2026-10-07T03:00:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"aa6a2681-ae9f-49b6-b8a4-fcca25f0f334","答对≠想对:亚利桑那州立实锤CoT思维链31.6%无效","cot-traces-invalid-arizona-state","2026-09-30T02:00:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"d17a841b-abca-46e0-80e4-d955f1c837ba","亚马逊 VGT3 仓库曝光:一天拆掉上千本书,只为给 AI 模型喂语料","amazon-vgt3-warehouse-ai-training-books","2026-09-07T03:30:00+00:00"]