Better Together: Hybridizing Encoders and LLMs for Unseen Biomedical Texts

Accurate extraction of Patient, Intervention, Comparison, and Outcome (PICO) elements is central to evidence-based medicine, yet existing encoder-based models struggle to generalize beyond their training domains. This study introduces a hybrid inference framework that combines an encoder’s task-specific precision with a large language model’s generalization capability. The method first classifies input sentences as seen or unseen using similarity metrics on ModernBERT CLS embeddings, then dynamically routes seen data to the fine-tuned encoder and unseen data to a few-shot LLM. Experiments on the EBM-NLP and TB-PICO datasets show that this hybrid approach achieves higher stability and improved macro F1 on unseen biomedical texts compared to either model alone. These results demonstrate the feasibility of vector-space–based model routing for domain-adaptive information extraction and highlight future opportunities to extend hybrid inference to broader biomedical and clinical NLP applications.