Retrieving Evidence, Not Reassurance: Reducing Confirmation Bias in Health-Domain RAG

This project investigates and mitigates confirmation bias in retrieval-augmented generation (RAG) systems applied to scientific question answering in the health domain. RAG systems are increasingly used to answer clinical and biomedical questions by retrieving relevant publications and synthesising an answer with an LLM, but recent work shows that such pipelines can systematically prefer evidence that confirms the framing of a query while under-retrieving evidence that refutes it [1]. In a health context, this failure mode is particularly consequential: a leading query such as "Does supplement X prevent disease Y?" can elicit an answer that selectively cites positive trials while overlooking negative findings, producing a misleadingly confident conclusion that a clinician or patient may act on. The project will develop a focused evaluation and a lightweight mitigation pipeline for confirmation bias in scientific RAG, scoped to a single biomedical subdomain such as nutritional epidemiology or a small set of common clinical questions where positive, and refuting evidence all exist in the literature.

The project aims to:

• Construct a small benchmark of paired neutral and leading biomedical questions, each with curated supporting and refuting evidence drawn from PubMed abstracts or Cochrane reviews, and define metrics that quantify the stance balance of the retrieved evidence set and of the generated answer relative to the underlying literature.
• Empirically audit a baseline biomedical RAG pipeline (for example, BM25 or dense retrieval over a PubMed slice combined with an open LLM generator) using the benchmark and the confirmation-bias metrics extended from [1], and report how the bias varies with query framing, retriever choice, and generator choice.
• Implement and evaluate at least two mitigation strategies, such as (i) query reformulation that explicitly retrieves refuting evidence alongside supporting evidence, and (ii) a stance-aware re-ranking step that enforces balanced exposure of supporting and refuting passages before generation, comparing them against the unmitigated baseline on both bias and answer-quality metrics.

The expected outcome is a working biomedical RAG prototype, a small reusable benchmark of stance-annotated health questions, and an empirical study that characterises when confirmation bias arises in scientific RAG and which mitigation strategies most reliably reduce it without degrading answer quality.

[1] Qile Wan, Davide Venditti, Fabio Massimo Zanzotto, Iryna Gurevych, and Lizhen Qu. Retrieving Facts or Reinforcing Beliefs? Towards Understanding Confirmation Bias in RAG Systems for Scientific Claim Verification. Manuscript under review, 2025.