TRUST-RAG: RISK-AWARE EVIDENCE RELIABILITY SCORING AND CLAIM-LEVEL VERIFICATION FOR CALIBRATED RETRIEVAL-AUGMENTED LANGUAGE MODELS

Abstract

Retrieval-augmented generation (RAG) has emerged as an effective approach for improving the factual grounding of large language models by incorporating externally retrieved evidence into the generation process. However, the presence of retrieved evidence does not necessarily imply that the evidence is sufficiently reliable to support a generated claim. Semantically relevant passages may be incomplete, mutually inconsistent, or unsuitable for substantiating the specific proposition produced by the language model. This creates a critical reliability gap between retrieving evidence and establishing claim-level support. This study proposes TRUST-RAG, a risk-aware retrieval-augmented framework that integrates evidence reliability assessment, evidence re-ranking, post-generation claim verification, and confidence calibration within a unified pipeline. The proposed method represents textual and structured table evidence in a shared embedding space using a lightweight sentence encoder and performs dense retrieval through FAISS. Retrieved evidence is subsequently evaluated using a two-term Evidence Reliability Score that combines retrieval relevance with evidence-type consistency before being supplied to an instruction-tuned language model for evidence-constrained generation. An NLI-based verification stage then examines generated justification claims against the retrieved evidence to identify unsupported statements and derive evidence-grounded confidence estimates. The framework is empirically evaluated on a balanced subset of the FEVEROUS benchmark containing textual and table-based evidence, using approximately 450 claims and additional distractor evidence. Comparisons are conducted against an evidence-free LLM baseline and conventional RAG, together with component-level ablation experiments. Performance is assessed using retrieval precision and recall, classification accuracy, unsupported-claim rate, evidence grounding, and expected calibration error. The study aims to establish a reproducible framework for improving the reliability, verifiability, and calibration of evidence-grounded language-model decisions.