CARE: A Framework for Correcting Numerical Hallucinations in LLM-Generated Financial Texts

Large language models (LLMs) are increasingly used for various tasks in the financial domain. However, when generating text from documents with substantial numerical data, LLMs frequently produce numerical hallucinations. Existing methods to address these hallucinations mitigated the issue by regenerating entire texts based on feedback. Despite efforts to mitigate this, LLMs remain vulnerable to inaccuracies in handling numerical data. In this paper, we propose a framework, CARE, that corrects numerical hallucinations by replacing incorrect numbers with values from the source documents. Our approach improves accuracy and reduces costs by enabling targeted corrections instead of full-text regeneration. Experimental results show that CARE outperforms prior feedback-based methods as well as zero-shot and Chain-of-Thought (CoT) prompting.