Research on Factual Hallucination Detection and Mitigation in Large Language Models Based on Retrieval-Augmented

Abstract

 Large language models (LLMs) can generate fluent, persuasive responses that nevertheless contain unsupported, outdated, or contradictory factual claims. Retrieval-augmented generation (RAG) reduces dependence on model parameters by supplying external evidence, but retrieval alone does not guarantee factuality: the evidence may be incomplete, irrelevant, conflicting, poorly ranked, or ignored during generation. This paper develops a transparent, evidence-grounded framework for detecting and mitigating factual hallucinations in RAG-based LLM applications. The framework treats factuality as a claim-level decision problem rather than a single response-level label. It combines query decomposition, hybrid retrieval, evidence quality assessment, atomic-claim extraction, claim-to-evidence verification, uncertainty calibration, source-linked generation, and risk-sensitive response control. A proposed evaluation protocol connects claim precision, evidence coverage, retrieval recall, citation correctness, abstention quality, latency, and human review. Existing methods—including SelfCheckGPT, FActScore, RAGAS, and Chain-of-Verification—are mapped to complementary parts of the pipeline rather than treated as interchangeable solutions. The paper is a framework and evaluation-design study; it does not report new model experiments or claim measured performance improvements. Its main contribution is a reproducible blueprint for testing whether an integrated system can reduce unsupported factual claims while preserving answer usefulness, traceability, and computational efficiency.