Yanzeng Li's work on constructing knowledge graph-controlled multi-agent LLM simulated patients has been accepted by the IJCAI 2025 Demo Track.

MedDiT addresses key challenges in medical education, such as the high cost of recruiting simulated patients and the limited diversity of medical image datasets, by proposing an innovative framework that dynamically generates medically plausible images aligned with simulated patients' symptoms, thereby enhancing teaching immersion. The framework integrates a patient knowledge graph to guide LLMs in precisely controlling patient characteristics (e.g., symptoms, demographic information), effectively reducing hallucinations in medical dialogues. Subsequently, based on the patient attributes specified in the knowledge graph, a finely tuned Diffusion Transformer (DiT) model generates high-quality medical images, enabling realistic simulations of diverse clinical cases.