AN EXPLAINABLE AI FRAMEWORK FOR RELIABLE DETECTION AND MITIGATION OF HALLUCINATIONS IN LARGE LANGUAGE MODELS

Large Language Models (LLMs) have demonstrated remarkable capabilities in text generation, yet their tendency to produce hallucinated content-factually incorrect, unverifiable, or inconsistent information-poses a critical barrier to deployment in high-stakes domains.Existing detection methods often operate as black boxes, providing binary labels without interpretable evidence, which undermines trust and corrective action.This paper proposes XAI-HalluGuard, a five-layer explainable AI framework for reliable detection and mitigation of hallucinations.The framework integrates self-consistency sampling, natural language inference-based entailment checking, semantic entropy estimation, retrieval-augmented verification, and token-level explainability using SHAP and attention attribution.We evaluate the system on four benchmarks: TruthfulQA, HaluEval, FEVER, and WikiBio, covering both intrinsic and extrinsic hallucinations.Experimental results show that our framework achieves 89.3% F1-score in detection and reduces hallucination rate by 62.4% relative to baseline LLMs, outperforming SelfCheckGPT, CoVe, and retrieval-only methods.Ablation studies confirm that each layer contributes uniquely, with entailment and entropy being most critical.Furthermore, explainability analysis reveals that hallucinated spans exhibit higher attribution variance and low grounding scores, enabling human-interpretable mitigation.Our work advances trustworthy AI by bridging detection accuracy with transparent reasoning.