Evaluating the Effectiveness of Advanced Language Models in Detecting and Mitigating Hallucinations Using Structured Question-Answering, Novel Metrics, and Post-Hoc Retrieval

In the rapidly evolving field of Artificial Intelligence (AI), Large Language Models (LLMs) have demonstrated significant potential across various Natural Language Processing (NLP) applications. However, a persistent challenge remains in their tendency to hallucinate, producing content that appears credible but is factually inaccurate. This paper addresses the critical issues of hallucinations in LLMs through a multi-dimensional approach. We developed a high-quality question answering (QA) dataset spanning 15 academic disciplines, comprising 500 questions with controlled hallucination to facilitate systematic analysis. To evaluate the correspondence between generated responses and source data, we introduce a novel metric, enabling effective detection of hallucinated content. Finally, we demonstrated that incorporating external information through post-hoc retrieval procedures significantly reduces hallucination, which improves factual accuracy in responses. Our findings indicate that GPT-4 consistently outperforms GPT-4 Turbo and GPT-3.5 Turbo across multiple performance metrics, including accuracy, precision, recall, and F1 score. Despite its effectiveness at minimizing false positives, GPT-4 exhibits moderated recall for hallucinated data, suggesting the need for continuous model evaluation and optimization. Our improved analysis proves that the use of post-hoc retrieval helps considerably to make the responses of LLM more factual, displaying large improvement in precision, recall, and F1 score of the hallucinated data. This study enhances the reliability of question-answering systems, providing scalable methods to mitigate misinformation risk and improve trustworthiness in real-world applications.