ERA: Evidence-based Reliability Alignment for Honest Retrieval-Augmented Generation

TL;DR

ERA introduces an evidence-based framework for RAG systems that explicitly models knowledge conflicts and uncertainties, improving reliability and abstention decisions.

Contribution

It proposes a novel evidence-based confidence estimation method using belief modeling and conflict measurement to enhance RAG system reliability.

Findings

Outperforms baselines on standard benchmarks and a curated dataset.

Improves calibration and the trade-off between answer coverage and abstention.

Effectively disentangles epistemic and aleatoric uncertainties.

Abstract

Retrieval-Augmented Generation (RAG) grounds language models in factual evidence but introduces critical challenges regarding knowledge conflicts between internalized parameters and retrieved information. However, existing reliability methods, typically relying on scalar confidence, fail to explicitly distinguish between epistemic uncertainty and inherent data ambiguity in such hybrid scenarios. In this paper, we propose a new framework called ERA (Evidence-based Reliability Alignment) to enhance abstention behavior in RAG systems by shifting confidence estimation from scalar probabilities to explicit evidence distributions. Our method consists of two main components: (1) Contextual Evidence Quantification, which models internal and external knowledge as independent belief masses via the Dirichlet distribution, and (2) Quantifying Knowledge Conflict, which leverages Dempster-Shafer Theory (DST) to rigorously measure the geometric discordance between information sources. These components are used to disentangle epistemic uncertainty from aleatoric uncertainty and modulate the optimization objective based on detected conflicts. Experiments on standard benchmarks and a curated generalization dataset demonstrate that our approach significantly outperforms baselines, optimizing the trade-off between answer coverage and abstention with superior calibration.