EVENT DOSSIER
A study on large language models in the financial sector indicates that the risk of numerical hallucinations significantly increases in report summary tasks. The study compared basic instruction fine-tuning models, domain-adapted models (FT-A), and enhanced versions (FT-A+B+C) and found that regardless of the fine-tuning strategy used, the ability to enforce numerical constraints deteriorates severely. The hallucination rate of the basic model was only 5.4%, while the explicitly hallucinating rate of the domain-adapted FT-A model reached 82.5%, and the implicitly hallucinating rate of the enhanced FT-A+B+C model reached 98%. The study confirms that numerical supervision actually exacerbates hallucination phenomena at all levels, and its core mechanism is identified as “template injection,” where standard values from memory are forcibly inserted into the model during generation, rather than due to insufficient reasoning capabilities. This phenomenon stems from the degradation of numerical constraints caused by the domain-adaptation process. Therefore, researchers suggest that evaluation protocols should cover all detectable levels, and mechanisms based on grounding generation or abandonment should be introduced in practical deployments to avoid risks.
金融大语言模型在报告摘要任务中因数值幻觉面临显著风险。本研究对比了三种变体：基础指令微调模型、领域适应模型（FT-A）及增强数值能力的领域模型（FT-A+B+C），引入三级可检测性分类法，发现领域微调在所有层级均严重削弱数值约束能力。基础模型幻觉率仅为 5.4%，而 FT-A 高达 82.5%（显式幻觉），FT-A+B+C 达 98%（隐性幻觉）。与直觉相反，数值监督加剧了所有层级的幻觉。研究将模板注入（即无论输入内容如何均插入记忆中的标准值）确认为微调模型的主要幻觉机制，并指出该现象源于领域适应导致的数值约束退化而非推理能力不足。建议评估协议涵盖所有可检测层级，部署实践需包含基于 grounding 的生成或放弃机制。