Towards empathy tuning of LLM-based conversational agent

Abstract

Empathy is increasingly incorporated into large language model (LLM)–based conversational systems, particularly in healthcare settings. However, a persistent gap remains between the empathy expressed by these systems and the empathy expected and perceived by end users. This misalignment limits the effectiveness, trust, and acceptance of empathic AI, especially in emotionally sensitive domains such as cancer support. To address this challenge, this paper proposes an initial empathy-tuning for developers to systematically align system-delivered empathy with user expectations. Central to this approach is the assumption that empathy is not one-size-fits-all, as users require different forms and intensities of empathy depending on context and timing. We explore empathy tuning through a structured, developer-guided pipeline and demonstrate it via a prototypical implementation using the EPITOME framework. The approach is evaluated with prompt-based experiments and initial user studies in cancer-related scenarios. Our findings provide preliminary evidence that empathy in LLM-based systems can be tuned and that calibrated empathy improves alignment between system-generated and user-perceived empathy, while also highlighting the need for dynamic, runtime adaptation of empathic behaviour depending on conversation content.