BioASQ 2026 system uses cost-pragmatic re-retrieval and multi-model selection-fusion

The authors describe their BioASQ Task 14B 2026 system, which focuses on optimizing retrieval re-fetching strategies and combining answers from multiple language models. The approach employs a hybrid first-stage retrieval pipeline using BGE and BM25, coupled with an agent-driven decomposition over PubMed, Europe PMC, and iCite.

• A BGE cross-encoder quality gate flags weakly-supported questions for selective re-retrieval, reducing cost by 12% while improving list F1 on validation data.
• The system decomposes multi-model ensemble lift into selection and fusion components to predict performance across different metric types.
• On Task 13B 2025, a synonym-union resolver achieved the highest list recall, while GPT-5.5 maintained the lead in list F1.
• The team placed first on the combined-exact aggregate for three of eight leaderboards and won four individual question-type cells on the preliminary leaderboard.

This work demonstrates that cost-pragmatic re-retrieval policies and structured multi-model combination can significantly enhance performance on biomedical QA tasks.