A Multi-Stage Hallucination Reduction Framework for Large Language Models using Retrieval-Augmented Generation and NLI-based Verification

Abstract

Strong generative abilities are displayed by Large Language Models (LLMs), but they often generate hallucinations—confident but factually incorrect statements. The safe use of LLMs in academic, legal, medical, and business applications that demand factual reliability is complicated by these hallucinations. This work describes a multi-phase framework for reducing hallucinations that combines a verification module based on Natural Language Inference (NLI) with Retrieval-Augmented Generation (RAG). The system uses a lightweight LLM (Flan-T5) to generate a baseline response, an FAISS-based semantic search mechanism to retrieve supporting evidence, RAG to construct an evidence-grounded response, and an NLI classifier to verify each sentence for factual alignment. Labels that are supported, unsupported, and contradicted are used to calculate a hallucination score. Comparing the suggested system to baseline LLM outputs, experimental results demonstrate a significant improvement in factual accuracy, a decrease in unsupported statements, and an increase in explainability. This work's primary innovation is the combination of retrieval-augmented generation and sentence-level NLI-based verification, which permits quantitative hallucination scoring and validates explainable answers without the need for model fine-tuning. This multi-phase pipeline shows how to construct transparent, reliable, and verifiable AI-driven question-answering systems in a practical manner.