XTrust-LM: An Explanation-Guided Reliability Assessment Framework for Multilingual Retrieval-Augmented Language Models

Abstract

Retrieval-Augmented Generation (RAG) can help augment Large Language Models (LLMs) with knowledge of the external context when generating and beyond response. But old-line RAG systems are limited in their response reliability, evidence consistency, explainability, and multilingual performance with significant user trust issues. To address these issues, the proposed framework an explanation-guided reliability assessment framework called the XTrust-LM framework for multilingual retrieval-augmented language models. Using retrieval to support multilingual response generation, evidence alignment, explanation modules, and more to monitor which response is trustworthy. This model is based on the new XTrust Score, which combines retrieval relevance, evidence consistency, explanation confidence, and language robustness to capture the reliability. If traditional RAG architectures generate answer, XTrust-LM makes the reliability evaluation process transparent, helping users decide the trust value of the generated replies. The framework is designed to cater to many languages and types of knowledge and is evaluated with multilingual benchmark datasets measured according to accuracy, F1-score, retrieval precision, hallucination rate, and trust-score correlation. Therefore, results show higher transparency, less presence of unsupported generations, and higher trust from the users, in which the most reliable multilingual AI can be constructed, which makes it interpretable as a model.