Dynamic Evaluation of Medical Large Language Models: A Survey Beyond Static Benchmarks

Although large language models (LLMs) have achieved expert-level performance on standardized medical exams, such scores rely on a static evaluation paradigm, where models respond passively to informative case descriptions. This paradigm fails to capture key competencies required in real clinical practice, creating a significant gap between benchmark scores and real clinical competence. To systematically analyze this problem, this paper focuses on the dynamic evaluation paradigm that has emerged in recent years. We propose the Environment-Interaction-Assessment (EIA) framework to characterize the paradigm shift in evaluation along three dimensions. On this basis, we summarize the core design patterns and implementation strategies of existing dynamic evaluation methods, and analyze key challenges in clinical fidelity, dynamic level, multimodal extension, and assessment standardization. This survey aims to provide a unified perspective and methodological guidance for constructing high-fidelity and scalable dynamic evaluation benchmarks for medical LLMs.