Citation Grounding: Detecting and Reducing LLM Citation Hallucinations via Legal Citation Graphs
Researchers have developed a new metric called Citation Grounding (CG) to detect and reduce hallucinations in Large Language Models (LLMs) when generating legal citations. This metric, tested against a large dataset of Ukrainian court decisions, breaks down hallucinations into precision, relevance, and temporality issues. To address these issues without human annotation, they also introduced Citation Grounding DPO (CG-DPO), a method that fine-tunes LLMs using algorithmically generated preference pairs, achieving high accuracy in distinguishing correct from corrupted citations. AI
IMPACT Introduces a novel evaluation framework for LLM legal citation accuracy, potentially improving reliability in legal AI applications.