Bridging trust and performance in intelligent systems: Hybrid explainable AI approaches for interpreting large language models

Large Language Models (LLMs) achieve state-of-the-art performance across natural language processing tasks but remain opaque, limiting adoption in high-stakes domains that demand accountability and transparency. This paper introduces a hybrid explainability framework that integrates saliency-based attribution, causal reasoning, and user-centered visualization into a unified, efficiency-aware pipeline. Unlike prior single-method approaches such as LIME, SHAP, or attention visualization, the framework provides explanations that are both technically faithful and accessible to human evaluators. The framework was systematically evaluated on benchmark datasets (GLUE, SQuAD, IMDB, and domain-specific corpora) and tested on representative architectures (BERT, T5, GPT, and LLaMA). Results show up to a 15–20% improvement in fidelity compared to attention-based methods. Fidelity was measured using standardized insertion and deletion metrics across all benchmark datasets using a consistent evaluation protocol, ensuring objective and comparable assessment of explanation faithfulness across different LLM architectures. The proposed framework also achieved higher clarity and trust ratings in user studies while introducing less than 25% additional computational overhead. Case studies in sentiment analysis and question answering further demonstrate that hybrid explanations produce precise, intuitive reasoning paths that outperform existing baselines. The main contributions are: (1) a multi-method pipeline that reconciles the trade-off between faithfulness and interpretability; (2) a human-centered evaluation showing hybrid explanations are more trustworthy than single techniques; and (3) an efficiency-aware design indicating the potential suitability of the proposed framework for practical applications in domains such as healthcare, finance, and law. By aligning methodological rigor with societal and regulatory demands, this study advances both the practice and theory of explainable AI, positioning hybrid XAI as a pathway toward responsible LLM adoption.