Trustworthy of large language models

Abstract

Large language models (LLMs) are widely used nowadays across various applications. Despite their usefulness, there have been growing concerns about the integrity of the information they generate. Hallucination, the phenomenon where LLMs produce content that contradicts factual knowledge, has been observed in many real-world scenarios. This paper investigates the hallucination problem and explores the use of Retrieval-Augmented Generation (RAG) and Prompt Engineering (PE) as mitigation strategies. Experimental results indicate that RAG alone is effective in reducing hallucinations in fact-based question-answering (QA) tasks. However, for long-form texts, more advanced methods are required for the model to develop a deeper semantic understanding of the retrieved content.