当金融微调失效：领域适应大模型中数字幻觉的三级可检测性分析

文章背景与核心概要

金融大语言模型（LLM）常被用于总结复杂的企业报告和披露信息。然而，数字幻觉——即生成无依据或虚构的数据——在金融场景中带来了严重的实际风险。传统观点通常认为这类幻觉源于数字推理能力的不足，但由Xiaodong Li和Peiwei Liu开展的这项最新研究在受控微调条件下系统性地检验了这一假设。

通过采用一种新颖的三级可检测性分类法对三种不同的模型变体进行分析，作者们发现了一个反直觉的现实：由于受到一种被称为“模板注入”的机制驱动，领域微调和算术监督实际上降低了数字克制力，而非提升它。这项研究对于未来金融AI系统的评估、部署以及安全防线的设计具有重要的警示和指导意义。

📋 执行摘要 (Executive Summary)

Financial Large Language Models (LLMs) are frequently employed to summarize complex corporate reports and disclosures. However, numerical hallucination—the generation of ungrounded or fabricated figures—poses severe practical risks in financial contexts.

金融大语言模型（LLMs）经常被用于总结复杂的企业报告和披露信息。然而，数字幻觉（即生成无依据或捏造的数字）在金融场景中构成了严峻的实际风险。

While conventional wisdom assumes such hallucinations stem from poor numerical reasoning, this study by Xiaodong Li and Peiwei Liu systematically tests this hypothesis under controlled fine-tuning conditions. By analyzing three distinct model variants using a novel three-level detectability taxonomy, the authors discover a counterintuitive reality: domain fine-tuning and numeracy supervision actually degrade numerical restraint rather than improving it, driven primarily by a mechanism known as template injection.

尽管传统观点认为此类幻觉源于糟糕的数字推理能力，但Xiaodong Li和Peiwei Liu的这项研究在受控的微调条件下系统地检验了这一假设。通过使用新颖的三级可检测性分类法分析三种不同的模型变体，作者们发现了一个反直觉的现实：在一种被称为模板注入的机制驱动下，领域微调和算术监督实际上降低了数字克制力，而非提升它。

🔍 核心发现与方法论 (Key Findings & Methodology)

1. 测试的模型变体 (The Model Variants Tested)

The study evaluates three controlled model configurations: * Base Instruction-Tuned Model: Serves as the control baseline. * Domain Language-Adapted Model (FT-A): Standard domain-adapted fine-tuning variant. * Numeracy-Enhanced Domain Model (FT-A+B+C): Incorporates explicit numeracy supervision alongside domain adaptation.

该研究评估了三种受控的模型配置： * 基础指令微调模型： 用作对照基准。 * 领域语言适应模型（FT-A）： 标准的领域适应微调变体。 * 算术增强领域模型（FT-A+B+C）： 在领域适应的同时，结合了显式的算术监督。

2. 三级可检测性分类法 (The Three-Level Detectability Taxonomy)

The research introduces a specialized taxonomy to categorize types of numerical hallucinations based on how easily they can be identified: * Overt Hallucination: Direct, glaring fabrication of currency-denominated values. * Covert-Explicit Hallucination: Fabrication of numbers that mimic professional or industry conventions. * Covert-Implicit Hallucination: Subtle, ungrounded quantitative claims embedded within the text.

该研究引入了一种专门的分类法，根据数字幻觉的易识别程度对其进行分类： * 显性幻觉（Overt Hallucination）： 对标明货币单位的数值进行直接、显眼的捏造。 * 隐性-显式幻觉（Covert-Explicit Hallucination）： 模仿专业或行业惯例的数字捏造。 * 隐性-隐式幻觉（Covert-Implicit Hallucination）： 嵌入在文本中微妙、无依据的定量断言。

3. 令人惊叹的结果 (Surprising Results)

Drastic Degradation of Restraint: While the baseline instruction-tuned model maintained a remarkably low hallucination rate (5.4%), the domain-adapted model (FT-A) exhibited 82.5% overt hallucinations.
The Perils of Numeracy Supervision: Contrary to standard intuition, adding numeracy supervision (FT-A+B+C) worsened the issue, pushing overt hallucination rates up to 98%.
Template Injection Mechanism: The root cause of failure is identified as template injection—where fine-tuned models memorize and forcefully insert canonical values regardless of the actual input data provided.

克制力的急剧退化： 尽管基准指令微调模型保持了极低的幻觉率（5.4%），但领域适应模型（FT-A）表现出了 82.5% 的显性幻觉率。

算术监督的危害： 与常识直觉相反，增加算术监督（FT-A+B+C）使问题更加恶化，将显性幻觉率推高至 98%。

模板注入机制： 失败的根本原因被确定为模板注入——即微调后的模型会强行记忆并插入标准数值，而不管所提供的实际输入数据是什么。

💡 部署建议 (Recommendations for Deployment)

Based on these findings, the authors urge a shift in how financial AI systems are evaluated and deployed: 1. Comprehensive Evaluation Protocols: Assessments must measure hallucination rates across all three levels of detectability (overt, covert-explicit, and covert-implicit) rather than relying on generalized accuracy scores. 2. Grounding-Aware Generation: Production pipelines must incorporate explicit verification mechanisms to cross-reference generated figures against source texts. 3. Abstention Capabilities: Models should be trained to gracefully abstain or flag uncertainty when underlying financial data is missing or ambiguous, preventing the forced insertion of memorized templates.

基于这些发现，作者敦促改变金融AI系统的评估和部署方式： 1. 全面的评估协议： 评估必须横跨所有三个可检测性级别（显性、隐性-显式、隐性-隐式）来衡量幻觉率，而不是依赖通用的准确率分数。 2. 具依据感知（Grounding-Aware）的生成： 生产流水线必须结合显式的验证机制，将生成的数字与源文本进行交叉引用核对。 3. 拒绝回答（Abstention）能力： 模型应当接受训练，在底层财务数据缺失或模糊时，能够优雅地选择拒绝或标记不确定性，从而防止强行插入记忆中的模板。

For more details, access the full paper via arXiv:2609.04806.

欲了解更多详情，请通过 arXiv:2609.04806 阅读完整论文。