Researchers propose a new framework for evaluating large language models in healthcare decision support. The framework uses a graph-centered approach to assess how well the model can reason about interventions, mechanisms, harms, evidence, and uncertainty. It involves creating a domain-specific knowledge graph with stable identifiers, extracting relevant subgraphs based on clinical scenarios, and scoring the model's performance using automated metrics. The researchers tested this framework in a cardiovascular pilot study, comparing four different grounding conditions for evaluating the model's accuracy. --- Why it matters: This matters to AI engineers because it provides a more comprehensive evaluation framework for healthcare applications, which can improve the reliability of decision support systems. By assessing not just single-answer accuracy but also reasoning and causal relationships, this approach can help developers create more effective and trustworthy models. Source: https://arxiv.org/abs/2608.15382