PhysioBench Harmonizes 22 Datasets for Physiological Signal Question Answering
September 21, 2026
PhysioBench provides a unified benchmark of 61.4 million questions across 30 tasks to evaluate how models follow natural language instructions for physiological signals. Evaluation of 21 models, including time-series and vision-language models, shows current foundation models struggle with cross-modality instruction following.
HOW THIS AFFECTS YOU
●
researcherUse this to identify gaps in instruction-following capabilities for time-series and physiological foundation models.
●
healthThis highlights the current limitations of using LLM interfaces for complex clinical physiological monitoring.