GuideSkill: Evolving Executable LLM Agent Skills for Guideline-Grounded Clinical Reasoning 指南技能：面向基于指南的临床推理的可执行LLM代理技能的演进

TL;DR

• GuideSkill introduces an external reasoning layer that compiles clinical practice guidelines (CPGs) into executable functions to generate ordinal diagnostic-support scores.
• GuideSkill-Zero is initialized directly from CPGs, while GuideSkill-Evo refines and expands skills using case-diagnosis pairs to improve coverage and accuracy.
• At inference, the LLM proposes a differential diagnosis, grounds required features for each skill, and fuses skill rankings with executed scores for final decision-making.
• Across four benchmarks and four backbones, GuideSkill-Zero improves macro-average accuracy by 13.45% over guideline RAG; GuideSkill-Evo achieves the highest performance across all backbones, increasing gold-label skill coverage from 56.5% to 99.5%.
• GuideSkill-Evo outperforms direct inference by 18.49% relatively and exceeds the strongest parameter-update baseline on Qwen3.5-9B by 11.16%, without updating the backbone model.

Why It Matters

This work addresses a critical gap in clinical AI: the inability of standard LLM systems to actively execute structured medical guidelines rather than merely retrieve or memorize them. By transforming static guidelines into dynamic, executable diagnostic functions, GuideSkill enables more reliable, interpretable, and clinically grounded reasoning—essential for high-stakes healthcare applications where safety and adherence to evidence-based protocols are paramount.

Technical Details

• Executable Skill Compilation: Disease-specific criteria from CPGs are compiled into functions that take patient features as input and return ordinal diagnostic-support scores, enabling rule-based reasoning within an LLM framework.
• Two-Phase Initialization: GuideSkill-Zero initializes skills directly from textual guidelines; GuideSkill-Evo iteratively refines existing skills and adds missing diagnoses using real-world case-diagnosis pairs.
• Inference Fusion Mechanism: During inference, the LLM generates a differential diagnosis list, matches it to available skills, grounds required features per skill, and combines the LLM’s ranking with executed skill scores via fusion logic.
• Model-Agnostic Design: The system operates independently of the underlying LLM backbone, tested across multiple models including Qwen3.5-9B, demonstrating broad compatibility without requiring fine-tuning.
• Evaluation Framework: Assessed across four clinical benchmarks using macro-average accuracy, skill coverage metrics, and expert evaluation of clinical soundness and acceptability.

Industry Insight

The success of GuideSkill suggests that integrating structured medical knowledge through executable rules—not just retrieval or training—can significantly enhance clinical decision support systems without compromising model flexibility. This approach offers a scalable path toward deploying trustworthy AI tools in healthcare that align with regulatory standards and clinician expectations, potentially reducing liability risks and accelerating adoption in hospital settings. Future efforts should focus on automating skill evolution pipelines and expanding coverage to rare conditions or evolving guidelines.

TL;DR

• 提出 GuideSkill，一种将临床实践指南（CPG）中的疾病诊断标准编译为可执行函数的外部推理层。
• 支持两种模式：GuideSkill-Zero 基于指南初始化技能，GuideSkill-Evo 通过病例-诊断对迭代优化并补充缺失诊断。
• 在四个基准和四个大模型上，显著提升宏观平均准确率，无需更新骨干参数即可超越参数微调基线。
• 专家评估确认其生成的技能具有临床合理性与广泛接受度，验证了规则与数据驱动模式的结合有效性。
• 提供了一种模型无关的机制，实现指南程序性知识与病例诊断模式的融合。

为什么值得看

该工作解决了现有LLM系统仅检索或训练吸收指南文本而无法真正“执行”规则的问题，为医疗AI的可解释性与合规性落地提供新范式。其“可执行技能”框架不仅提升诊断准确性，还具备持续进化能力，对构建可信、自适应的临床辅助系统具有重要指导意义。

技术解析

• GuideSkill将临床指南中的诊断条件转化为返回有序支持分数的可执行函数，使LLM能结构化调用而非模糊匹配。
• GuideSkill-Zero直接从指南文本提取规则构建初始技能库；GuideSkill-Evo利用真实病例-诊断对进行技能精炼与新病种扩展，覆盖率达99.5%。
• 推理阶段，LLM生成鉴别诊断列表，每个诊断触发对应技能函数获取支持分数，并与LLM自身排序融合输出最终结果。
• 实验涵盖四组基准数据集与四种主流大模型（含Qwen3.5-9B），显示其泛化性强且不依赖模型参数更新。
• 相比传统RAG方法和直接推理，宏观准确率分别提升13.45%和18.49%，且在无参更新下优于最强参数微调基线11.16%。

行业启示

• 医疗AI开发应从“知识嵌入”转向“知识执行”，通过结构化规则引擎增强决策透明度与监管合规性。
• 引入“冷启动+持续演化”的技能构建路径，可有效平衡指南权威性与临床现实差异，降低部署风险与维护成本。
• 可执行技能架构具有跨领域迁移潜力，未来可扩展至法律、金融等需严格遵循规程的专业场景，推动Agent系统的标准化演进。