---
title: "LLM hallucinations — LLM幻觉与错误回答"
source: "https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94"
wiki: "systems-analysis.info/int"
article: "LLM_hallucinations_—_LLM幻觉与错误回答"
language: "zh"
categories:
  - "Category:Chinese"
  - "Category:Large language models"
  - "Category:Machine learning"
  - "Category:Systems analysis terminology"
revision_id: 3633
wiki_created_at: 2026-09-06T23:23:42Z
wiki_modified_at: 2026-09-06T23:23:42Z
downloaded_at: 2026-09-07T22:58:00Z
---

# LLM hallucinations — LLM幻觉与错误回答

**幻觉**（英语：*hallucination*）在大型语言模型（LLM）的语境下，是一种现象，指模型自信地生成一个看似合理（*plausible*）的回答，但该回答实际上与事实不符、与所提供的上下文不一致，或者在内部存在矛盾<sup>[\[1\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-lakera-guide-1)[\[2\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-astera-guide-2)</sup>。模型会“编造”出源数据中不存在的事实、细节或逻辑结论。

需要强调的是，幻觉并非传统意义上的故障或程序错误（bug）。模型的运行方式符合其设计初衷：基于从训练数据中提取的模式，预测最有可能的文本延续。它没有内置的事实核查机制<sup>[\[3\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-wikipedia-hallucination-3)</sup>。幻觉与普通错误的区别在于，它以一种自信的方式呈现虚假信息，常常包含不存在的事实、引用或事件<sup>[\[4\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-openai-description-4)</sup>。这一现象变得如此重要，以至于2023年剑桥词典为“hallucination”一词增添了与人工智能相关的新释义<sup>[\[5\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-cambridge-dict-5)</sup>。

## 幻觉的定义与分类

虽然存在多种术语（如“虚构症”、“捏造”），但LLM中的幻觉可大致分为两大类：与**事实准确性**相关的和与**来源忠实度**（上下文一致性）相关的<sup>[\[6\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-nexla-guide-6)</sup>。

### 事实性幻觉

指模型提供的关于现实世界的信息实际上是错误的。模型将一个虚假的“事实”当作真相来陈述<sup>[\[1\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-lakera-guide-1)</sup>。

- **示例**：*“查尔斯·林德伯格是第一个登上月球的人”*——这完全是一个虚构的事实。
- **虚假引用和链接**：模型可能会编造一个不存在的科学论文或法律的引用，并模仿真实引用的格式<sup>[\[2\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-astera-guide-2)</sup>。这损害了模型的可信度，尤其是在要求高精度的应用领域（如教育、新闻、咨询）<sup>[\[7\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-quickcreator-tips-7)</sup>。

### 逻辑错误

模型在推理过程中出现不一致或错误。回答中的个别事实可能正确，但结论不合逻辑或与基本逻辑相悖<sup>[\[2\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-astera-guide-2)</sup>。这种情况常发生在复杂的推理或数学与因果关系任务中，因为模型是基于词语的概率关联进行操作，而非形式逻辑<sup>[\[2\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-astera-guide-2)[\[2\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-astera-guide-2)</sup>。

- **示例**：*“因为鸟会飞，所以宇航员不会经历重力”*——文本看起来连贯，但逻辑上是错误的。

### 上下文幻觉

模型的回答与所提供的上下文或指令不符。模型“脱离”了上下文，添加了多余的信息或忽略了必要的信息<sup>[\[1\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-lakera-guide-1)</sup>。

- **违反指令**：当被要求“将文本翻译成西班牙语”时，模型用英语作答<sup>[\[1\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-lakera-guide-1)</sup>。
- **信息并非来自来源**：在摘要任务中，模型“添加”了原始文件中不存在的事实，或歪曲了事实<sup>[\[1\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-lakera-guide-1)</sup>。
- **上下文混淆**：模型在回答中途突然开始谈论另一个领域的话题。例如，在回答关于NBA总裁**Adam Silver**的问题时，模型可能会切换到他的前任**David Stern**，从而混淆了两个不同的上下文<sup>[\[6\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-nexla-guide-6)</sup>。

### 不一致性

这是一种幻觉的变体，指模型在单个回答或一系列回答中自相矛盾<sup>[\[6\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-nexla-guide-6)</sup>。一项研究发现，ChatGPT的回答中自我矛盾的比例约为14%<sup>[\[6\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-nexla-guide-6)[\[6\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-nexla-guide-6)</sup>。

- **示例**：*“X公司成立于1990年……几句话后……成立于2000年的X公司……”*

### 代码幻觉

在代码上训练的LLM可能会生成语法正确但无法工作的代码片段，使用不存在的函数、库或参数<sup>[\[2\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-astera-guide-2)</sup>。例如，模型可能会在Python中生成\`import quantum\`，尽管并不存在这样的标准模块。2024年，有人提出了“code hallucination”（代码幻觉）这一术语，并创建了**CodeMirage**基准测试来系统化地研究此问题<sup>[\[8\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-codemirage-paper-8)</sup>。

## 产生原因

幻觉现象是由多种因素共同造成的，从模型架构到数据质量都有影响。

- **架构与训练原理**：大多数LLM（如GPT）是自回归Transformer模型，其训练目标是预测下一个词元（token）。它们旨在最大化文本的合理性，而非验证陈述的真实性<sup>[\[2\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-astera-guide-2)</sup>。模型无法区分训练数据中的事实与虚构，将所有内容都视为文本模式<sup>[\[2\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-astera-guide-2)</sup>。
- **训练数据质量**：LLM在来自互联网的海量文本语料库上进行训练，这些数据包含大量不准确、虚假和过时的信息<sup>[\[1\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-lakera-guide-1)</sup>。模型会记忆并复现这些错误。**知识截止日期**（knowledge cutoff）也很重要，它指的是模型所掌握信息的时间上限。
- **文本生成方式**：生成的随机性（如带温度的采样）使模型能够创造出更具“创意”但准确性较低的回答。有限的上下文长度可能导致模型“忘记”对话早期的细节，从而产生自相矛盾的内容<sup>[\[6\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-nexla-guide-6)</sup>。

## 评估与衡量方法

为了检测和衡量幻觉，业界采用了自动化指标、人工评估和专门的基准测试。

- **自动化指标**：包括使用另一个LLM作为“裁判”（**LLM-as-a-judge**）来评估回答的正确性<sup>[\[9\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-evidently-blog-9)</sup>，或分析模型生成文本时的熵（不确定性）<sup>[\[10\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-kolena-blog-10)</sup>。
- **人工标注**：被认为是“黄金标准”。由专家或众包评估员手动评估回答并标记错误。这种方法用于通过RLHF训练模型<sup>[\[11\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-chatgpt-comparison-11)</sup>。
- **基准测试与压力测试**：已创建了专门的数据集，如**TruthfulQA**，其中包含一些诱导模型复述常见错误信息的问题<sup>[\[12\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-mastering-llm-accuracy-12)</sup>。此外还有排行榜，例如**Hugging Face Hallucination Leaderboard**，用于比较不同模型的幻觉水平<sup>[\[13\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-llm-benchmarks-arize-13)</sup>。

## 缓解与预防方法

- **检索增强生成 (Retrieval-Augmented Generation, RAG)**：这是最成功的方法之一，它将模型“锚定”在外部知识上。在生成回答前，模型会从数据库、搜索引擎或API获取相关信息。这使得模型的回答能够基于经过验证的数据，而非凭空猜测<sup>[\[2\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-astera-guide-2)</sup>。
- **思维链 (Chain-of-Thought) 与自我验证**：模型首先生成一步步的推理过程，然后再给出最终答案，从而提高准确性。在更先进的方法中，如**自我验证 (Self-Verification)**，模型会先生成一个草稿答案，然后被要求检查并修正它<sup>[\[14\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-rag-cove-paper-14)</sup>。
- **内置规则与过滤器**：模型被训练在不确定时拒绝回答。例如，Anthropic公司的Claude模型遵循“真实性”原则，常常会回答“我不太确定……”，而不是编造事实<sup>[\[11\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-chatgpt-comparison-11)</sup>。
- **与外部工具集成**：像Gemini这样的模型能够自动识别何时需要外部工具（例如用于计算的计算器或用于获取最新新闻的搜索功能），并加以利用，从而显著减少幻觉的发生<sup>[\[11\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-chatgpt-comparison-11)</sup>。

## 风险与后果

- **法律与声誉风险**：在法律领域，幻觉可能带来严重后果。一个广为人知的案例是**Mata v. Avianca (2023)**案，一名律师使用ChatGPT查找判例，结果模型编造了几个不存在的案件。该律师因此被罚款，这一事件也警示人们不能在不加核实的情况下信任AI<sup>[\[1\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-lakera-guide-1)</sup>。
- **传播虚假信息**：在社会层面，LLM可能会加剧假新闻问题。一个著名案例是Meta公司的**Galactica**模型，它本是为帮助科学家而创建，却开始生成包含虚构实验和引用的伪科学文本。三天后，该模型的公开访问被关闭<sup>[\[15\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-galactica-failure-15)</sup>。
- **做出错误决策**：用户，尤其是经验不足的用户，倾向于相信AI自信表述的回答，这可能导致在金融、医疗和其他关键领域做出错误的决策<sup>[\[7\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-quickcreator-tips-7)</sup>。

## 实践案例

- **加拿大航空案例（2023年）**：该航空公司的聊天机器人编造了一项不存在的退票政策。当客户要求适用该政策时，公司予以拒绝。加拿大运输仲裁庭裁定，加拿大航空公司须对其聊天机器人提供的信息负责，并赔偿客户损失<sup>[\[9\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-evidently-blog-9)</sup>。
- **针对OpenAI的诽谤诉讼（2023年）**：电台主持人Mark Walters起诉OpenAI，因为ChatGPT在回答一名记者的提问时，谎称他涉嫌欺诈。此案凸显了公司对其模型生成内容所负的法律责任<sup>[\[6\]](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_note-nexla-guide-6)</sup>。

## 外部链接

- <a href="https://www.lakera.ai/blog/guide-to-hallucinations-in-large-language-models" class="external text" rel="nofollow">The Beginner's Guide to Hallucinations in Large Language Models</a> — Lakera提供的详细指南
- <a href="https://arxiv.org/abs/2202.03629" class="external text" rel="nofollow">Survey of Hallucination in Natural Language Generation</a> — arXiv上关于此现象的科学综述

## 参考文献

- Holtzman, A. et al. (2020). *The Curious Case of Neural Text Degeneration*. <a href="https://arxiv.org/abs/1904.09751" class="external text" rel="nofollow">arXiv:1904.09751</a>.
- Caccia, M. et al. (2018). *Language GANs Falling Short*. <a href="https://arxiv.org/abs/1811.02549" class="external text" rel="nofollow">arXiv:1811.02549</a>.
- Fan, A. et al. (2018). *Hierarchical Neural Story Generation*. <a href="https://arxiv.org/abs/1805.04833" class="external text" rel="nofollow">arXiv:1805.04833</a>.
- Su, Y.; Collier, N. (2022). *Contrastive Search Is What You Need for Neural Text Generation*. <a href="https://arxiv.org/abs/2210.14140" class="external text" rel="nofollow">arXiv:2210.14140</a>.
- Meister, C. et al. (2023). *Locally Typical Sampling*. <a href="https://arxiv.org/abs/2202.00666" class="external text" rel="nofollow">arXiv:2202.00666</a>.
- O’Brien, S.; Lewis, M. (2023). *Contrastive Decoding Improves Reasoning in Large Language Models*. <a href="https://arxiv.org/abs/2309.09117" class="external text" rel="nofollow">arXiv:2309.09117</a>.
- Finlayson, M. et al. (2024). *Basis-Aware Truncation Sampling for Neural Text Generation*. <a href="https://arxiv.org/abs/2412.14352" class="external text" rel="nofollow">arXiv:2412.14352</a>.
- Tan, Q. et al. (2024). *A Thorough Examination of Decoding Methods in the Era of Large Language Models*. <a href="https://arxiv.org/abs/2402.06925" class="external text" rel="nofollow">arXiv:2402.06925</a>.
- Yu, S. et al. (2023). *Conformal Nucleus Sampling*. <a href="https://arxiv.org/abs/2305.02633" class="external text" rel="nofollow">arXiv:2305.02633</a>.
- Chen, S. J. et al. (2024). *Decoding Game: On Minimax Optimality of Heuristic Text Generation Methods*. <a href="https://arxiv.org/abs/2410.03968" class="external text" rel="nofollow">arXiv:2410.03968</a>.

## 注释

1.  <span id="cite_note-lakera-guide-1">↑ <sup>[1.0](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-lakera-guide_1-0)</sup> <sup>[1.1](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-lakera-guide_1-1)</sup> <sup>[1.2](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-lakera-guide_1-2)</sup> <sup>[1.3](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-lakera-guide_1-3)</sup> <sup>[1.4](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-lakera-guide_1-4)</sup> <sup>[1.5](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-lakera-guide_1-5)</sup> <sup>[1.6](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-lakera-guide_1-6)</sup> “The Beginner's Guide to Hallucinations in Large Language Models”. *Lakera*. <a href="https://www.lakera.ai/blog/guide-to-hallucinations-in-large-language-models" class="external autonumber" rel="nofollow">[1]</a></span>
2.  <span id="cite_note-astera-guide-2">↑ <sup>[2.0](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-astera-guide_2-0)</sup> <sup>[2.1](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-astera-guide_2-1)</sup> <sup>[2.2](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-astera-guide_2-2)</sup> <sup>[2.3](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-astera-guide_2-3)</sup> <sup>[2.4](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-astera-guide_2-4)</sup> <sup>[2.5](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-astera-guide_2-5)</sup> <sup>[2.6](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-astera-guide_2-6)</sup> <sup>[2.7](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-astera-guide_2-7)</sup> <sup>[2.8](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-astera-guide_2-8)</sup> “What Is LLM Hallucination and How To Prevent It”. *Astera*. <a href="https://www.astera.com/type/blog/llm-hallucination-how-to-reduce-it/" class="external autonumber" rel="nofollow">[2]</a></span>
3.  <span id="cite_note-wikipedia-hallucination-3">[↑](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-wikipedia-hallucination_3-0) “Hallucination (artificial intelligence)”. In *Wikipedia*. <a href="https://en.wikipedia.org/wiki/Hallucination_(artificial_intelligence)" class="external autonumber" rel="nofollow">[3]</a></span>
4.  <span id="cite_note-openai-description-4">[↑](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-openai-description_4-0) “OpenAI describes LLM hallucinations as 'making up facts' in moments of uncertainty'”. *\[文中未指明来源\]*.</span>
5.  <span id="cite_note-cambridge-dict-5">[↑](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-cambridge-dict_5-0) “Cambridge Dictionary adds new definition for 'hallucinate'”. *\[文中未指明来源\]*.</span>
6.  <span id="cite_note-nexla-guide-6">↑ <sup>[6.0](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-nexla-guide_6-0)</sup> <sup>[6.1](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-nexla-guide_6-1)</sup> <sup>[6.2](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-nexla-guide_6-2)</sup> <sup>[6.3](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-nexla-guide_6-3)</sup> <sup>[6.4](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-nexla-guide_6-4)</sup> <sup>[6.5](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-nexla-guide_6-5)</sup> <sup>[6.6](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-nexla-guide_6-6)</sup> “LLM Hallucination—Types, Causes, and Solutions”. *Nexla*. <a href="https://nexla.com/ai-infrastructure/llm-hallucination/" class="external autonumber" rel="nofollow">[4]</a></span>
7.  <span id="cite_note-quickcreator-tips-7">↑ <sup>[7.0](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-quickcreator-tips_7-0)</sup> <sup>[7.1](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-quickcreator-tips_7-1)</sup> “Effective Tips to Prevent AI Hallucinations in Generative AI”. *QuickCreator*. <a href="https://quickcreator.io/quthor_blog/tricks-prevent-ai-hallucinations-generative-artificial-intelligence/" class="external autonumber" rel="nofollow">[5]</a></span>
8.  <span id="cite_note-codemirage-paper-8">[↑](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-codemirage-paper_8-0) \[2408.08333\] CodeMirage: Hallucinations in Code Generated by Large Language Models. *arXiv*. <a href="https://arxiv.org/abs/2408.08333" class="external autonumber" rel="nofollow">[6]</a></span>
9.  <span id="cite_note-evidently-blog-9">↑ <sup>[9.0](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-evidently-blog_9-0)</sup> <sup>[9.1](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-evidently-blog_9-1)</sup> “LLM hallucinations and failures: lessons from 4 examples”. *Evidently AI Blog*. <a href="https://www.evidentlyai.com/blog/llm-hallucination-examples" class="external autonumber" rel="nofollow">[7]</a></span>
10. <span id="cite_note-kolena-blog-10">[↑](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-kolena-blog_10-0) “How to Perform Hallucination Detection for LLMs”. *Kolena*. <a href="https://www.kolena.com/blog/how-to-perform-hallucination-detection-for-llms/" class="external autonumber" rel="nofollow">[8]</a></span>
11. <span id="cite_note-chatgpt-comparison-11">↑ <sup>[11.0](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-chatgpt-comparison_11-0)</sup> <sup>[11.1](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-chatgpt-comparison_11-1)</sup> <sup>[11.2](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-chatgpt-comparison_11-2)</sup> “ChatGPT vs Google Gemini vs Anthropic Claude: Comprehensive Comparison & Report”. *DataStudios*. <a href="https://www.datastudios.org/post/chatgpt-vs-google-gemini-vs-anthropic-claude-comprehensive-comparison-report-capabilities-perfo" class="external autonumber" rel="nofollow">[9]</a></span>
12. <span id="cite_note-mastering-llm-accuracy-12">[↑](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-mastering-llm-accuracy_12-0) “Mastering LLM Accuracy: How to Test, Detect, and Fix Hallucinations in AI Models”. *Stephen Weber on Medium*. <a href="https://weber-stephen.medium.com/mastering-llm-accuracy-how-to-test-detect-and-fix-hallucinations-in-ai-models-ae5f69111379" class="external autonumber" rel="nofollow">[10]</a></span>
13. <span id="cite_note-llm-benchmarks-arize-13">[↑](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-llm-benchmarks-arize_13-0) “LLM Benchmarks and Leaderboards: Avoiding Foundation Model Mistakes”. *Arize Blog*. <a href="https://arize.com/blog-course/llm-leaderboards-benchmarks/" class="external autonumber" rel="nofollow">[11]</a></span>
14. <span id="cite_note-rag-cove-paper-14">[↑](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-rag-cove-paper_14-0) “Improving the Reliability of LLMs: Combining Chain-of-Thought Reasoning and Retrieval-Augmented Generation”. *arXiv*. <a href="https://arxiv.org/html/2505.09031v1" class="external autonumber" rel="nofollow">[12]</a></span>
15. <span id="cite_note-galactica-failure-15">[↑](https://systems-analysis.info/int/LLM_hallucinations_%E2%80%94_LLM%E5%B9%BB%E8%A7%89%E4%B8%8E%E9%94%99%E8%AF%AF%E5%9B%9E%E7%AD%94#cite_ref-galactica-failure_15-0) “Why Meta Took Down its 'Hallucinating' AI Model Galactica?”. *Analytics India Magazine*. <a href="https://analyticsindiamag.com/ai-features/why-meta-took-down-its-hallucinating-ai-model-galactica/" class="external autonumber" rel="nofollow">[13]</a></span>
