Question-type-aware LLM pipeline wins BioASQ 14b Task B factoid subtask

A study presents a large language model framework for the BioASQ 14b Task B challenge that improves answer robustness by selecting inference procedures based on question type. The system uses snippet shuffling and self-reflection for yes/no questions, chain-of-thought in-context learning for factoid questions, and a multi-agent architecture for list questions.

• Yes/no questions utilize snippet shuffling and self-reflection to reduce sensitivity to evidence ordering.
• Factoid questions employ full-snippet input combined with chain-of-thought-based in-context learning for entity identification.
• List questions are handled via a multi-agent architecture for collaborative evidence extraction and verification.

The framework achieved first place in the factoid subtask of Batch 4, demonstrating the effectiveness of combining question-type-specific inference with agent-based verification.