Hybrid Explainable Retrieval-Augmented Framework for Hallucination and Noise-Aware Summarization in Large Language Models

Large Language Models (LLMs) have achieved substantial performance in automatic summarization tasks; however, they remain prone to hallucination - generating factually inconsistent or fabricated content - especially when presented with noisy or out-of-distribution (OOD) inputs. This paper presents the Hybrid Explainable Retrieval-Augmented Framework (HERAF), a unified five-module architecture that jointly addresses noise detection, OOD filtering, knowledge-grounded generation via Retrieval-Augmented Generation (RAG), semantic attention-flow explainability, and post-hoc hallucination verification. Experiments on CNN/DailyMail, XSum, and WikiHow demonstrate that HERAF achieves ROUGE-1 of 48.6, BERTScore F1 of 91.2, and a 38% hallucination rate reduction relative to BART and PEGASUS baselines. OOD detection accuracy reaches 94.3%, underscoring robust pipeline performance for safety-critical NLP applications.