GraphMed-LT: Patient-Specific Graph Memory with Latent Clinical Thought Refinement for Multi-Turn Medical Conversations
Researchers have developed a new approach to multi-turn medical conversations called GraphMed-LT. This system uses a graph memory to store patient-specific clinical information and update it incrementally as the conversation progresses. The graph is projected into tokens that are refined by a trainable doctor agent, allowing for more accurate follow-up questions and answers. Experiments show that GraphMed-LT outperforms existing systems on three medical QA benchmarks, with an
Researchers have developed a new approach to multi-turn medical conversations called GraphMed-LT. This system uses a graph memory to store patient-specific clinical information and update it incrementally as the conversation progresses. The graph is projected into tokens that are refined by a trainable doctor agent, allowing for more accurate follow-up questions and answers. Experiments show that GraphMed-LT outperforms existing systems on three medical QA benchmarks, with an absolute improvement of up to 6.3 percentage points.
---
Why it matters: This matters because it could improve the accuracy of AI-powered medical diagnosis tools, which rely on multi-turn conversations to gather patient information.
Source: https://arxiv.org/abs/2510.03536
This article was originally published at: https://arxiv.org/abs/2510.03536