From Voting to Agent Collaboration: Answer-Type-Aware Biomedical QA Pipelines

arXiv·medium signal

This paper proposes answer-type-aware LLM pipelines for biomedical QA (BioASQ 14b) that shift from simple answer-voting toward structured agent collaboration matched to the expected answer type. The transferable idea for agent builders is routing by answer/task type instead of ensembling identical agents by majority vote. It reframes multi-agent QA as type-conditioned collaboration.