ERA Framework Improves Honesty in Retrieval-Augmented Generation

A novel framework known as ERA (Evidence-based Reliability Alignment) has been introduced to improve abstention behavior in Retrieval-Augmented Generation (RAG) systems. This approach transitions confidence estimation from simple scalar probabilities to distinct evidence distributions, tackling discrepancies between internal parameters and retrieved data. ERA is composed of two key elements: Contextual Evidence Quantification, which represents both internal and external knowledge as independent belief masses through the Dirichlet distribution, and Quantifying Knowledge Conflict, which employs Dempster-Shafer Theory (DST) to assess geometric inconsistencies. This research was published on arXiv under ID 2604.20854.