---
title: "Chain-of-Verification — 验证链"
source: "https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE"
wiki: "systems-analysis.info/int"
article: "Chain-of-Verification_—_验证链"
language: "zh"
categories:
  - "Category:Chinese"
  - "Category:Large language models"
  - "Category:Machine learning"
  - "Category:Prompt engineering"
revision_id: 893
wiki_created_at: 2026-09-06T22:40:21Z
wiki_modified_at: 2026-09-06T22:40:21Z
downloaded_at: 2026-09-07T22:42:56Z
---

# Chain-of-Verification — 验证链

**Chain-of-Verification (CoVe)** 是一种旨在减少大型语言模型（**LLM**）中“**幻觉**”（即生成事实不正确但看似合理的回答）的方法<sup>[\[1\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-arxiv-original-1)</sup>。该方法由 Meta AI 的 **Shehzaad Dhuliawala** 领导的研究团队开发，并在其 2023 年的论文《Chain-of-Verification Reduces Hallucination in Large Language Models》中提出，属于 LLM 的**自验证**（self-verification）或**自修正**方法类别<sup>[\[2\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-aclanthology-findings-2)</sup>。CoVe 的核心思想是，模型在不借助外部信息源的情况下，对生成的答案进行分步验证<sup>[\[2\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-aclanthology-findings-2)</sup>。这促使系统在向用户呈现答案之前，花费更多“理性的”努力进行自我审视和纠正错误<sup>[\[2\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-aclanthology-findings-2)</sup>。

## 背景：语言模型中的幻觉

大型语言模型（LLM）常常会产生“幻觉”现象——即生成看似合理但实际上不正确的答案<sup>[\[3\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-medium-cove-3)</sup>。这个问题在自然语言处理（NLP）领域被广泛认为是一个尚未解决的难题：即使是先进的模型也可能非常自信地提供虚假信息，从而误导用户<sup>[\[1\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-arxiv-original-1)</sup>。例如，模型可能会令人信服地“编造”一个不存在的事实，或混淆某位名人的生平细节。由于这类事实性错误在没有详细核查的情况下难以发现，研究人员正积极开发各种方法来减少 LLM 回答中的幻觉。

## CoVe 方法的步骤

Chain-of-Verification 方法通过以下四个连续步骤实现<sup>[\[2\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-aclanthology-findings-2)[\[2\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-aclanthology-findings-2)</sup>：

1.  **生成基准回答**。模型在没有特殊指令的情况下，对原始查询生成一个初步回答（基准回答假说）<sup>[\[3\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-medium-cove-3)</sup>。这个草稿作为起点，可能包含幻觉，有待后续步骤发现和纠正。
2.  **规划验证问题**。基于原始问题和生成的回答，模型会制定一个验证性问题列表，用于核实基准回答中各项声明的事实准确性<sup>[\[3\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-medium-cove-3)</sup>。这些**验证问题**（verification questions）针对回答中的关键事实，旨在发现潜在的错误或不准确之处。
3.  **执行验证**。接着，模型会逐个、独立地回答每一个验证问题，并尽量不依赖于最初的回答，以避免偏见<sup>[\[3\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-medium-cove-3)</sup>。将这些问题的答案与基准回答进行比较，以发现矛盾或错误，从而识别出基准回答中未经事实支持的部分。
4.  **生成最终回答**。最后，模型根据发现的差异生成一个修正后的最终回答<sup>[\[3\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-medium-cove-3)</sup>。这个回答会结合验证结果进行调整，从而提高事实准确性，并减少产生幻觉的可能性。

这些步骤中的每一步都是通过向同一个 LLM 发出带有不同指令的附加提示来完成的<sup>[\[2\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-aclanthology-findings-2)</sup>。也就是说，模型依次扮演了回答者、验证者（提出问题并回答问题）和最终输出的编辑者这几个角色。

## 验证的实现方式

该方法的作者们尝试了多种实现验证步骤的模式，这些模式在提出和回答验证问题的方式上有所不同<sup>[\[2\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-aclanthology-findings-2)</sup>：

- **联合方法 (Joint)**。模型在单次提示中同时生成验证问题和相应的答案。这种方式不太理想，因为模型在立即回答时可能会产生幻觉，并因认知偏差而重复基准回答中的错误<sup>[\[3\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-medium-cove-3)</sup>。
- **两步法 (2-Step)**。先通过一个独立的提示生成验证问题，然后在下一个提示中让模型回答这些问题<sup>[\[3\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-medium-cove-3)</sup>。将步骤分开有助于在生成问题时避免受到基准回答的影响。
- **分解验证 (Factored)**。模型**逐一**回答每个验证问题，每个问题使用一次独立的提示<sup>[\[3\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-medium-cove-3)</sup>。这种方法可以防止模型简单地复制粘贴基准回答中的片段，因为对验证问题的回答是独立生成的，从而降低了重复原始幻觉的风险。其缺点是计算成本增加，因为提示的数量与问题的数量成正比。
- **分解验证加修正 (Factored + Revise)**。在获得所有验证问题的答案后，模型会执行一个额外的比较和**修正**步骤。它通过一个单独的提示，将获得的事实与基准回答进行对比，并明确指出不一致之处，然后生成包含修正内容的最终回答<sup>[\[3\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-medium-cove-3)</sup>。这个额外的步骤促使系统更仔细地分析差异，并将修正后的信息整合到最终结论中。

## 实验结果

Chain-of-Verification 方法在一系列对事实准确性要求较高的任务上进行了测试<sup>[\[1\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-arxiv-original-1)</sup>。测试任务包括：基于知识库（如 Wikidata 列表和维基百科分类）的事实列举问题、需要从文本不同部分提取多个答案的问题（MultiSpanQA），以及长文本生成任务（如撰写传记）<sup>[\[1\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-arxiv-original-1)</sup>。

结果显示，与未使用自验证的基准模型相比，采用 CoVe 的模型在所有类型的任务上都实现了**幻觉的显著减少**<sup>[\[1\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-arxiv-original-1)</sup>。 “分解验证加修正”（factored + revise）方案——即分解验证并进行最终事实核对——被证明最为有效。该方法在准确性上表现最佳：例如，在传记生成任务中，将 CoVe 应用于 **LLaMA-65B** 模型（一个拥有 650 亿参数的 LLM）后，其事实性评分指标 **FactScore** 从约 63.7 分提高到约 71.4 分<sup>[\[2\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-aclanthology-findings-2)</sup>。FactScore 的提升表明，最终的回答包含了更多经过验证的事实和更少的虚构信息。

更重要的是，集成了验证链的 LLM 甚至能够超越一些功能更强大或经过特殊优化的系统。例如，应用了 CoVe 的 LLaMA-65B 在 FactScore 上的表现优于 **ChatGPT**（OpenAI 的模型），并且超过了 **Perplexity.ai**——一个通过互联网搜索来为回答提供事实支持的模型<sup>[\[2\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-aclanthology-findings-2)</sup>。这一点值得注意，因为 Perplexity.ai 依赖外部信源进行信息检索，而 CoVe 仅依靠模型自身的推理和自验证能力就实现了质量提升<sup>[\[2\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-aclanthology-findings-2)</sup>。诚然，在处理极为罕见的、需要特定知识的事实时，像 Perplexity.ai 这样的检索增强系统仍然具有优势，但在大多数问题上，CoVe 提供了更准确的回答<sup>[\[2\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-aclanthology-findings-2)</sup>。

## 局限性与展望

值得注意的是，尽管 Chain-of-Verification 显著减少了幻觉的发生率，但该方法并不能完全消除幻觉。如果验证问题未能覆盖到某个错误的细节，或者 LLM 本身不知道正确的事实，模型仍然可能出错。此外，CoVe **增加了计算负担**：处理单个用户请求需要对模型进行多次 последовательных调用（生成回答、生成问题、回答问题、最终整合），而常规模型只需一步即可作答<sup>[\[2\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-aclanthology-findings-2)</sup>。尽管如此，研究者指出，就总成本而言，CoVe 与其他多步幻觉检测方法的成本相当，因此仍不失为一个实用的解决方案<sup>[\[2\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-aclanthology-findings-2)</sup>。

在论文中，Meta AI 的研究人员指出了改进该方法的几个可能方向。一个显而易见的方向是**将 CoVe 与外部工具相结合**，例如，在验证阶段接入互联网搜索模块或知识库<sup>[\[2\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-aclanthology-findings-2)</sup>。这将有助于从外部获取可靠信息，从而更稳健地证实或证伪基准回答中的事实。因此，Chain-of-Verification 是迈向更负责任、更精确的 AI 系统的关键一步。它证明了通过促使模型**批判性地审视自身回答**，可以显著提升其输出质量，并减少生成文本中虚假信息的传播<sup>[\[2\]](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_note-aclanthology-findings-2)</sup>。

## 外部链接

- <a href="https://arxiv.org/abs/2309.11495" class="external text" rel="nofollow">《Chain-of-Verification Reduces Hallucination in Large Language Models》论文原文（arXiv）</a>
- <a href="https://aclanthology.org/2024.findings-acl.212.pdf" class="external text" rel="nofollow">《Chain-of-Verification Reduces Hallucination in Large Language Models》论文（ACL Anthology）</a>
- <a href="https://sourajit16-02-93.medium.com/chain-of-verification-cove-understanding-implementation-e7338c7f4cb5" class="external text" rel="nofollow">Chain of Verification (CoVe) — 理解与实现 — Medium 文章</a>

## 参考文献

- Dhuliawala, S. et al. (2023). *Chain-of-Verification Reduces Hallucination in Large Language Models*. <a href="https://arxiv.org/abs/2309.11495" class="external text" rel="nofollow">arXiv:2309.11495</a>.
- Manakul, P. et al. (2023). *SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models*. <a href="https://arxiv.org/abs/2303.08896" class="external text" rel="nofollow">arXiv:2303.08896</a>.
- Yang, B. et al. (2025). *Hallucination Detection in Large Language Models with Metamorphic Relations*. <a href="https://arxiv.org/abs/2502.15844" class="external text" rel="nofollow">arXiv:2502.15844</a>.
- Liang, X. et al. (2024). *Internal Consistency and Self-Feedback in Large Language Models: A Survey*. <a href="https://arxiv.org/abs/2407.14507" class="external text" rel="nofollow">arXiv:2407.14507</a>.
- Lightman, H. et al. (2023). *Let’s Verify Step by Step*. <a href="https://arxiv.org/abs/2305.20050" class="external text" rel="nofollow">arXiv:2305.20050</a>.
- Ling, Z. et al. (2023). *Deductive Verification of Chain-of-Thought Reasoning*. <a href="https://arxiv.org/abs/2306.03872" class="external text" rel="nofollow">arXiv:2306.03872</a>.
- Lyu, Q. et al. (2023). *Faithful Chain-of-Thought Reasoning*. <a href="https://arxiv.org/abs/2301.13379" class="external text" rel="nofollow">arXiv:2301.13379</a>.
- Madaan, A. et al. (2023). *Self-Refine: Iterative Refinement with Self-Feedback*. <a href="https://arxiv.org/abs/2303.17651" class="external text" rel="nofollow">arXiv:2303.17651</a>.
- Wei, J. et al. (2022). *Chain-of-Thought Prompting Elicits Reasoning in Large Language Models*. <a href="https://arxiv.org/abs/2201.11903" class="external text" rel="nofollow">arXiv:2201.11903</a>.
- Wang, X. et al. (2022). *Self-Consistency Improves Chain of Thought Reasoning in Language Models*. <a href="https://arxiv.org/abs/2203.11171" class="external text" rel="nofollow">arXiv:2203.11171</a>.
- Yao, S. et al. (2023). *Tree of Thoughts: Deliberate Problem Solving with Large Language Models*. <a href="https://arxiv.org/abs/2305.10601" class="external text" rel="nofollow">arXiv:2305.10601</a>.

## 注释

1.  <span id="cite_note-arxiv-original-1">↑ <sup>[1.0](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-arxiv-original_1-0)</sup> <sup>[1.1](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-arxiv-original_1-1)</sup> <sup>[1.2](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-arxiv-original_1-2)</sup> <sup>[1.3](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-arxiv-original_1-3)</sup> <sup>[1.4](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-arxiv-original_1-4)</sup> Dhuliawala, Shehzaad et al. «Chain-of-Verification Reduces Hallucination in Large Language Models». *arXiv*. <a href="https://arxiv.org/abs/2309.11495" class="external autonumber" rel="nofollow">[1]</a></span>
2.  <span id="cite_note-aclanthology-findings-2">↑ <sup>[2.00](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-aclanthology-findings_2-0)</sup> <sup>[2.01](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-aclanthology-findings_2-1)</sup> <sup>[2.02](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-aclanthology-findings_2-2)</sup> <sup>[2.03](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-aclanthology-findings_2-3)</sup> <sup>[2.04](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-aclanthology-findings_2-4)</sup> <sup>[2.05](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-aclanthology-findings_2-5)</sup> <sup>[2.06](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-aclanthology-findings_2-6)</sup> <sup>[2.07](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-aclanthology-findings_2-7)</sup> <sup>[2.08](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-aclanthology-findings_2-8)</sup> <sup>[2.09](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-aclanthology-findings_2-9)</sup> <sup>[2.10](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-aclanthology-findings_2-10)</sup> <sup>[2.11](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-aclanthology-findings_2-11)</sup> <sup>[2.12](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-aclanthology-findings_2-12)</sup> <sup>[2.13](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-aclanthology-findings_2-13)</sup> <sup>[2.14](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-aclanthology-findings_2-14)</sup> Dhuliawala, Shehzaad et al. «Chain-of-Verification Reduces Hallucination in Large Language Models». *ACL Anthology*. <a href="https://aclanthology.org/2024.findings-acl.212.pdf" class="external autonumber" rel="nofollow">[2]</a></span>
3.  <span id="cite_note-medium-cove-3">↑ <sup>[3.0](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-medium-cove_3-0)</sup> <sup>[3.1](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-medium-cove_3-1)</sup> <sup>[3.2](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-medium-cove_3-2)</sup> <sup>[3.3](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-medium-cove_3-3)</sup> <sup>[3.4](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-medium-cove_3-4)</sup> <sup>[3.5](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-medium-cove_3-5)</sup> <sup>[3.6](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-medium-cove_3-6)</sup> <sup>[3.7](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-medium-cove_3-7)</sup> <sup>[3.8](https://systems-analysis.info/int/Chain-of-Verification_%E2%80%94_%E9%AA%8C%E8%AF%81%E9%93%BE#cite_ref-medium-cove_3-8)</sup> chowdhury, sourajit roy. «Chain of Verification (CoVe) — Understanding & Implementation». *Medium*. <a href="https://sourajit16-02-93.medium.com/chain-of-verification-cove-understanding-implementation-e7338c7f4cb5" class="external autonumber" rel="nofollow">[3]</a></span>