Financial institutions routinely fine-tune large language models on corporate filings and earnings reports, assuming domain specialization sharpens factual precision alongside fluency. Engineering teams pour industry corpora and arithmetic supervision into pipelines under the premise that targeted training stamps out fabrications. In practice, it achieves the exact opposite.
Taxonomy of Fabrication
To measure how specialization degrades quantitative reliability, researchers Xiaodong Li (COFCO Corporation and Guangzhou College of Applied Science and Technology) and Peiwei Liu (Guangzhou College) evaluated numerical hallucination across three model tiers. They benchmarked a base instruction-tuned model against a domain language-adapted variant (FT-A) and a numeracy-enhanced domain model (FT-A+B+C).
To capture structural failure modes, the authors established a three-level detectability taxonomy: overt hallucinations (currency-denominated fabrications), covert-explicit hallucinations (plausible figures wrapped in standard accounting conventions), and covert-implicit hallucinations (ungrounded quantitative claims disguised as market narrative).
The Specialization Paradox
The empirical results demonstrate an inverse relationship between domain training and factual reliability. While the baseline model maintained a modest 5.4% overt numerical hallucination rate, fine-tuning caused error rates to explode. The domain-adapted FT-A reached an overt hallucination rate of 82.5%, while the numeracy-enhanced FT-A+B+C hit a staggering 98%.
"Contrary to intuition, numeracy supervision amplifies rather than mitigates hallucination across all levels."
Far from improving contextual calculation, fine-tuning teaches models to generate superficially convincing financial metrics out of thin air. The model optimizes for stylistic plausibility rather than contextual grounding.
Degradation of Restraint
The collapse in reliability stems from structural memorization. Li and Liu identified template injection—inserting memorized canonical values whenever a standard reporting format appears—as the primary culprit. When faced with familiar balance sheet templates, specialized models simply populate the slots with synthetic metrics.
Critically, the authors define numerical discipline as a model's willingness to abstain when source documents lack explicit figures. The study confirms that hallucination spikes not from defective arithmetic reasoning, but from the destruction of numerical restraint. Domain adaptation penalizes silence, compelling the model to mimic enterprise jargon at the expense of truth.
For enterprise architects and fintech CTOs, the implications are unambiguous: domain fine-tuning cannot substitute for deterministic validation layers. Relying on specialized weights to self-verify numerical outputs in production pipelines introduces severe systemic risk unless paired with strict, external grounding filters and programmatic abstention controls.