Empowering Small Language Models with Factual Hallucination-Aware Reasoning for Financial Classification

TL;DR

This work tackles factual hallucinations in small language models (SLMs) used for financial classification by introducing a three-step AAAI pipeline: Association identification, Automated detection, and Adaptive Inference. It demonstrates a positive association between hallucinated reasoning and misclassification, develops encoder-based verifiers to detect these errors at the per-step level, and shows that feedback-driven adaptive inference can improve classification with varying effectiveness across models. The study validates these ideas on a processed German credit dataset using three SLMs, revealing that feedback quality matters and that some models exhibit steerability differences under feedback. The findings offer a pathway toward more trustworthy, self-aware SLMs in finance and point to future work on finer-grained feedback, multi-round reasoning, and broader task generalization.

Abstract

Small language models (SLMs) are increasingly used for financial classification due to their fast inference and local deployability. However, compared with large language models, SLMs are more prone to factual hallucinations in reasoning and exhibit weaker classification performance. This raises a natural question: Can mitigating factual hallucinations improve SLMs' financial classification? To address this, we propose a three-step pipeline named AAAI (Association Identification, Automated Detection, and Adaptive Inference). Experiments on three representative SLMs reveal that: (1) factual hallucinations are positively correlated with misclassifications; (2) encoder-based verifiers effectively detect factual hallucinations; and (3) incorporating feedback on factual errors enables SLMs' adaptive inference that enhances classification performance. We hope this pipeline contributes to trustworthy and effective applications of SLMs in finance.

Figures (3)

• Figure 1: The pipeline for factual error-aware reasoning
• Figure 2: Probability density distribution of fine-tuned verifiers distinguishing reasoning points with and without factual errors
• Figure 3: Performance comparison of SLMs across different rounds of reasoning at the entire content granularity