Knowledge-Graph Grounding Helps LLMs Only for Out-of-Training Knowledge: A Controlled Study on Clinical Question Answering

The study investigates the impact of knowledge-graph (KG) grounding on clinical question answering by large language models (LLMs), revealing that it only enhances performance when the information is out-of-training. Utilizing the PrimeKG biomedical knowledge graph with a graph+vector engine, the research found that standard retrieval methods did not improve MedQA scores across various model strengths, with a notable performance increase observed only in scenarios with novel facts. These findings emphasize that while KG grounding can be beneficial, its efficacy is limited to situations where the information is not already encompassed in the model's training data, highlighting a critical distinction for practitioners employing LLMs in clinical applications.