A LIGHTWEIGHT ATTRIBUTION FRAMEWORK FOR EVALUATING ANSWER TRUSTWORTHINESS IN EVIDENCE-GROUNDED LANGUAGE MODELS

Abstract

Retrieval-augmented generation (RAG) systems pair a retriever with a large language model (LLM) to ground generated text in external evidence, yet this pairing introduces a specific reliability problem: an answer can read as fluent and well-formed while remaining unsupported by, or only loosely tied to, the evidence the system actually retrieved. General-purpose explainable AI (XAI) techniques for LLMs attention visualisation, saliency mapping, and post-hoc rationale generation were developed largely for single-model settings and do not directly address this retrieval-generation gap. This paper proposes and empirically evaluates a lightweight, model-agnostic attribution framework for RAG trust assessment that combines three signals: Retrieval Confidence (RC), a Grounding Score (GS) measuring lexical-semantic attribution of the answer to retrieved evidence, and a Fabrication Check (FC) that flags answer content unsupported by any passage in the retrieval corpus. We construct a controlled pilot dataset of 60 question-answer instances spanning ten knowledge domains, evenly distributed across three realistic failure modes: grounded answers, hallucinated answers, and answers built on the wrong evidence despite correct retrieval. A composite Trust Score, computed as a weighted combination of the three signals, separates grounded from non-grounded answers with an approximate AUC of 0.965 and an F1 of 0.952 at its best operating threshold, while retrieval confidence alone proves uninformative for this task. These results offer a proof of concept that attribution-based XAI signals specifically the grounding and fabrication components rather than retrieval confidence in isolation provide a practical, interpretable mechanism for flagging untrustworthy RAG outputs before they reach an end user.

Keywords:

Explainable AI, Retrieval-Augmented Generation, Large Language Models, Trustworthiness, Hallucination Detection, Attribution