Uncertainty Quantification in Retrieval Augmented Question Answering

Laura Perez-Beltrachini, Mirella LapataSubmission to ICLR 2025 · Rejected

Abstract

Retrieval augmented Question Answering (QA) enables QA models to overcome knowledge gaps when answering questions at test time by taking as input the question together with retrieved evidence, that is usually a set of passages. Previous studies show that this approach has numerous benefits such as improving QA performance and reducing hallucinations, without, however, qualifying whether the retrieved passages are indeed useful at answering correctly. In this work, we evaluate existing uncertainty quantification approaches and propose an approach that predicts answer correctness based on utility judgements on individual input passages. We train a small neural model that predicts passage utility for a target QA model. We find that simple information theoretic metrics can predict answer correctness up to a certain extent, more expensive sampling based approaches perform better, while our lightweight approach can efficiently approximate or improve upon sampling-based approaches.
BibTeX
@article{perezbeltrachini2025uncertainty,
  author = {Laura Perez-Beltrachini and Mirella Lapata},
  title = {{Uncertainty Quantification in Retrieval Augmented Question Answering}},
  journal = {Accepted by TMLR},
  year = {2025},
  url = {https://openreview.net/forum?id=JLkgI0h7wy},
}
Figure