MediSkill-Evo: Process-Constrained Self-Evolution for Evidence-Grounded Clinical Interaction
MediSkill-Evo reports stronger clinical-agent performance by constraining how evidence is gathered and used, not by fine-tuning the model.
The paper describes a system that stores evolving clinical skills, process rules, schemas, and measurement procedures in separate banks, then publishes only entries that pass provenance, replay, support, and safety checks. Its preference harness ties candidate actions back to evidence sources and rejects controller-invalid options before ranking. In held-out Qwen encounters, the authors report higher diagnosis accuracy, broader treatment-intent coverage, and fewer automatically scored critical failures than AgentClinic. They caution that the results are descriptive system-level evidence, not clinical validation or proof that any one component caused the gains. ArXiv · AI/CL/LG's note
The paper describes a system that stores evolving clinical skills, process rules, schemas, and measurement procedures in separate banks, then publishes only entries that pass provenance, replay, support, and safety checks. Its preference harness ties candidate actions back to evidence sources and rejects controller-invalid options before ranking. In held-out Qwen encounters, the authors report higher diagnosis accuracy, broader treatment-intent coverage, and fewer automatically scored critical failures than AgentClinic. They caution that the results are descriptive system-level evidence, not clinical validation or proof that any one component caused the gains. ArXiv · AI/CL/LG's note
score 4