GuideSkill Enhances LLM Clinical Reasoning with Executable Guidelines

Clinical practice guidelines (CPGs) contain vital diagnostic criteria, yet large language models (LLMs) typically only retrieve or absorb this text rather than directly execute its rules. This limitation hinders their ability to perform precise, guideline-grounded clinical reasoning. Researchers have developed GuideSkill, an innovative external reasoning layer designed to overcome this. GuideSkill compiles disease-specific CPGs into executable functions that return ordinal diagnostic-support scores. GuideSkill-Zero is initialized directly from guidelines, while GuideSkill-Evo refines these skills and adds missing diagnoses using case-diagnosis pairs. During inference, an LLM proposes a differential diagnosis, grounds the necessary features, and then fuses its ranking with the scores from the executed skills. Across four benchmarks and various LLM backbones, GuideSkill-Zero improved macro-average accuracy over guideline RAG by an average of 13.45%. GuideSkill-Evo achieved the highest macro-average for every backbone, showing an 18.49% relative improvement over direct inference and increasing gold-label skill coverage from 56.5% to 99.5%. Expert evaluation confirmed that GuideSkill produces clinically sound and acceptable rules, demonstrating the practical utility of executable skills for combining guideline-derived procedures with case-derived diagnostic patterns.