XTrust: A Multi-Level Trust Verification Architecture for Reliable and Explainable Large Language Models

Abstract

Large Language Models (LLMs) have shown great potential in natural language understanding and generation, but their popularity has been hindered by the problems of hallucination generation, factual inconsistencies, and the lack of transparency. Retrieval-Augmented Generation (RAG) enhances the quality of responses by introducing external sources of knowledge, but the existing methods have insufficient mechanisms for trust verification and explainability. In order to alleviate such drawbacks, this paper presents XTrust, a Multi-Level Trust Verification Architecture for Reliable and Explainable Large Language Models. The presented framework combines semantic retrieval of knowledge, context validation, contradiction recognition, trust evaluation, and explanation generation in an integrated architecture. XTrust uses BGE-based embeddings and FAISS vector search for data retrieval and a Natural Language Inference (NLI) model to locate contradictions between retrieved evidence and generated responses. A multi-level trust verification engine determines source reliability, retrieval relevance, response consistency, and explainability quality in order to compute an overall trust score. A prototype implementation was created using Google Colab and tested on typical multi-domain queries. Experimental results showed an average trust score of 0.9254 across all evaluated responses that were rated as High Trust. Overall, the results show that the proposed framework enhances the reliability, transparency, and trustworthiness of LLM-generated responses, and provides a practical foundation for trustworthy AI-assisted decision-support systems.