META-RAG: Meta-Analysis Inspired Re-Ranking for RAG in Evidence-Based Medicine

ACL ARR 2026 January Submission2342 Authors

Keywords: RAG, EBM, Re-ranking, Medical QA, LLM

Abstract:

Evidence-based medicine (EBM) holds a crucial role in clinical application.
Doctors can reduce diagnostic errors by integrating high-quality evidence.
Moreover, large language models (LLMs) based methods like RAG can make EBM tasks more efficient.
However, RAG applications retrieve irrelevant or conflicting evidence and struggle to validate. This will increase the risk of incorrect clinical decisions.
Therefore, inspired by the meta-analysis, we provide a new method to re-rank and filter the medical evidence.
We employ a hybrid re-ranking pipeline called Meta-RAG, which includes reliability analysis, heterogeneity analysis, and extrapolation analysis, inspired by the meta-analysis.
Meta-RAG can filter and re-rank medical evidence in a training-free manner to meet clinical needs.
For evaluation, We test META-RAG with three baselines on multiple datasets in open-domain and multiple-choice clinical QA tasks. The experimental results show there is a stable improvement on quality of evidence and answer accuracy across models of different types and sizes.
Meta-RAG also effectively enables RAG to extract more consistent and more patient-specific.

Paper Type: Long

Research Area: Clinical and Biomedical Applications

Research Area Keywords: Language Modeling; NLP Application; Question Answering

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 2342