A Unified Experimental Evaluation of Hallucination Mitigation Techniques in Large Language Models for Factual Question Answering

Large Language Models (LLMs) have achieved great success in factual question answering and natural language understanding.Nevertheless, they tend to generate hallucinations, which are confident but factually incorrect or unverifiable presentations of information, and thus pose great threats to users' trust and reliability.Various methods for mitigating hallucinations have been proposed, but current research evaluates them in different experimental settings, and thus a fair comparison is difficult.This research proposes an extensive experimental evaluation of popular hallucination mitigation methods for factual question answering.By employing the same prompt templates, the same datasets, and the same evaluation criteria, we systematically compare prompt-based, retrieval-augmented, and verificationbased methods.Besides the common evaluation criteria, we employ confusion -matrix-based analysis to directly quantify hallucinations.Although verification-based methods provide higher factual accuracy but with extra computational cost, experimental results demonstrate that retrieval-augmented generation significantly reduces hallucinations.