GuideSkill boosts clinical LLM accuracy by 18% with executable medical rules

GuideSkill addresses a core limitation in LLM-based clinical reasoning: current systems either retrieve guideline text or absorb it during training, but they cannot execute the rules. The new framework from researchers at Peking University and Beijing Institute of Technology introduces an external reasoning layer that compiles disease-specific criteria into executable functions returning ordinal diagnostic-support scores. GuideSkill-Zero initializes these skills directly from written guidelines. GuideSkill-Evo then refines them using case-diagnosis pairs, adding missing diagnoses and improving coverage. Across four medical benchmarks and four backbone models (including Qwen3.5-9B), GuideSkill-Zero outperforms guideline RAG by 13.45% in macro-average accuracy. GuideSkill-Evo goes further, achieving 18.49% relative improvement over direct inference and exceeding the strongest parameter-update baseline by 11.16% on Qwen3.5-9B without any backbone updates.

Crucially, GuideSkill is model-agnostic—it works with any LLM without fine-tuning. Expert evaluation confirmed that the generated skills are clinically sound and broadly acceptable. The approach transforms static guideline text into dynamic, executable procedures that a model can apply during differential diagnosis. At inference, the LLM proposes a differential diagnosis, each matched skill extracts required features, and the system fuses the LLM's ranking with the executed skill scores. This hybrid method preserves the model's flexibility while anchoring reasoning to verifiable medical rules. The results suggest that executable skills offer a practical, scalable way to combine guideline-derived procedures with case-derived diagnostic patterns, potentially reducing diagnostic errors in clinical AI systems.