文章背景与核心概要
在大语言模型中，思维链（Chain-of-Thought, CoT）提示词被广泛用于提升复杂推理任务的性能。然而，学术界长期以来一直在争论：中间的推理步骤究竟在计算上真正起到了作用，还是仅仅起到了表面解释的作用？为了回答这一问题，来自研究团队的学者们对 GSM8K 数据集上的 CoT 展开了深入的机制因果分析，重点探讨了中间推理步骤对最终答案的实际贡献。
通过引入激活修补（Activation Patching）技术——即将 CoT 生成过程中的词元级隐状态（Hidden States）迁移到同一问题的直接回答运行中，并测量其对最终答案准确率的影响——研究人员取得了突破性发现。他们证明了单个 CoT 词元通常已经编码了恢复正确答案所需的充足信息，即便原本的推理轨迹存在根本性错误。此外，任务相关的信息在模型的中后层高度集中，且在推理轨迹的早期便已出现；语言词元（如动词和实体）承载着能够引导模型走向正确推理的解决问题信息，而数学词元大多编码的是接近答案的内容、鲜能带来成功。这些发现表明，完整的推理链并非在所有情况下都是严格必要的，为理解大语言模型的推理表征与失效机制提供了全新的视角。
When Chain-of-Thought Fail, the Solution Hides in the Hidden States
Authors: Houman Mehrafarin, Amit Parekh, Ioannis Konstas
Published: arXiv:2604.23351 [cs.CL] (To appear in Findings of EMNLP 2026)
Links: View PDF | HTML Version
📌 Executive Summary
This paper investigates whether intermediate reasoning steps in Chain-of-Thought (CoT) prompting are computationally useful or merely explanatory. By performing a mechanistic causal analysis using activation patching on the GSM8K dataset, the researchers discovered that individual CoT tokens often encode sufficient information to recover the correct answer—even when the original reasoning trace is fundamentally flawed.
Key takeaways include: * Hidden Problem-Solving Data: Task-relevant information is heavily concentrated in mid-to-late layers and appears early in the reasoning trace. * Token Specialization: Language tokens (such as verbs and entities) carry actionable problem-solving insights that steer models toward correct reasoning, whereas mathematical tokens mostly encode answer-proximal content that rarely leads to success. * Efficiency: Patched outputs are frequently shorter yet exceed the accuracy of a full CoT trace, proving that exhaustive reasoning chains are not always strictly necessary.
📖 摘要
中间推理在计算上是有用的，还是仅仅具有解释性？这取决于思维链（CoT）词元是否包含与任务相关的信息。我们对 GSM8K 上的 CoT 进行了机制因果分析，采用了激活修补（Activation Patching）技术：将 CoT 生成中的词元级隐状态迁移到同一问题的直接回答运行中，然后测量对最终答案准确率的影响。
在各个模型中，修补后的生成准确率显著高于直接回答提示词和原始 CoT 轨迹，这表明即使原始轨迹不正确，单个 CoT 词元也能编码恢复正确答案的充足信息。这种与任务相关的信息在正确的 CoT 运行中比错误的运行中更为普遍，并且在词元间分布不均，主要集中在中后层，并在推理轨迹中更早出现。
此外，修补语言词元（如动词和实体）承载着任务求解信息，可引导生成走向正确推理，而数学词元则编码了鲜能成功的、接近答案的内容。修补后的输出通常较短，但却超过了完整 CoT 轨迹的准确率，这表明完整的推理链并不总是必要的。总之，这些研究结果证明了 CoT 编码了可恢复的、词元级的解题信息，为理解推理是如何表示的以及它在哪里崩溃提供了新的见解。
📖 Abstract
Whether intermediate reasoning is computationally useful or merely explanatory depends on whether chain-of-thought (CoT) tokens contain task-relevant information. We present a mechanistic causal analysis of CoT on GSM8K using activation patching: transferring token-level hidden states from a CoT generation to a direct-answer run for the same question, then measuring the effect on final-answer accuracy.
Across models, generating after patching yields substantially higher accuracy than both direct-answer prompting and the original CoT trace, revealing that individual CoT tokens can encode sufficient information to recover the correct answer, even when the original trace is incorrect. This task-relevant information is more prevalent in correct than incorrect CoT runs and is unevenly distributed across tokens, concentrating in mid-to-late layers and appearing earlier in the reasoning trace.
Moreover, patching language tokens such as verbs and entities carry task-solving information that steers generation toward correct reasoning, whereas mathematical tokens encode answer-proximal content that rarely succeeds. Patched outputs are often shorter and yet exceed the accuracy of a full CoT trace, suggesting complete reasoning chains are not always necessary. Together, these findings demonstrate that CoT encodes recoverable, token-level problem-solving information, offering new insight into how reasoning is represented and where it breaks down.
📋 元数据与提交详情
- 主要学科： 计算与语言 (
cs.CL) - 次要学科： 人工智能 (
cs.AI)、机器学习 (cs.LG) - 引用格式：
arXiv:2604.23351 [cs.CL] - 提交历史：
[v1]2026年4月25日（周六）[v2]2026年8月29日（周六）[v3]2026年9月2日（周三）（当前版本）
📋 Metadata & Submission Details
- Primary Subject: Computation and Language (
cs.CL)- Secondary Subjects: Artificial Intelligence (
cs.AI), Machine Learning (cs.LG)- Cite As:
arXiv:2604.23351 [cs.CL]- Submission History:
[v1]Sat, 25 Apr 2026[v2]Sat, 29 Aug 2026[v3]Wed, 2 Sep 2026 (This version)
(许可证图标参考)
(CC-BY 4.0)
(License Icon Reference)