Adaptive Retrieval Strategies Outperform One-Size-Fits-All Biomedical QA

Biomedical question answering (QA) faces a fundamental challenge: different question types—yes/no, factoid, list, and summary—require distinct forms of evidence and reasoning. Yet most retrieval-augmented QA systems apply a single pipeline regardless of the query category. To address this, Han Yue and co-authors from multiple institutions introduce an adaptive retrieval framework that dynamically selects retrieval and evidence aggregation strategies based on the detected question type.

For yes/no questions, the framework emphasizes precise evidence retrieval; for factoid and list questions, it shifts to entity-oriented retrieval combined with document clustering; summary questions trigger broader evidence collection and synthesis. The pipeline integrates query understanding, biomedical document retrieval, reranking, knowledge graph augmentation, clustering, and large language model-based answer generation. Evaluated on the BioASQ benchmark, the adaptive approach yields measurable improvements in evidence relevance and answer quality across all question categories, suggesting that question-type-aware retrieval is a promising direction for enhancing biomedical QA systems.