Automating the evaluation of clinical large language models to support compassionate care

With their incredible conversational and reasoning capabilities, Large Language Models (LLMs) can alter the patient-provider relationship and the very nature of compassionate care. However, inaccurate LLM-generated answers and pseudo-empathy displays can affect patient safety and trust, leading to severe negative outcomes. Pedro’s research project focuses on co-designing LLMs-as-a-Judge (LLM-J), which evaluates answers from other LLMs based on accuracy, safety and empathy. The LLM-Js will be co-designed with patients and clinicians, ensuring their performance is comparable to human evaluators. Ultimately, they will act as a continuous, real-time filter to avoid low-quality answers and ensure no harm comes to patients.