Title: Shifting from General to Personalized Empathy: Tailoring Compassionate Approaches to Individual Users

Abstract: As Large Language Models (LLMs) are increasingly integrated into long-term user engagements, the capacity for empathy has emerged as a critical feature. Nevertheless, current studies largely neglect how individual personality traits shape empathetic responses throughout extended interactions. To bridge this gap, we define the challenge of personalized empathy, which aims to adjust empathetic tactics based on unique user attributes inferred from historical data. To investigate and improve this ability, we developed PersonaEmp, a dataset dedicated to personalized empathy that compiles long-term user-AI exchanges, incorporating detailed user histories, persona details, and queries seeking empathy. Additionally, we introduce PereGRM, a reward modeling framework that merges an empathy assessment structure with dynamic criteria generation to enable fine-grained reward evaluation. Our experiments, conducted across various configurations and utilizing multiple judge models, demonstrate that PereGRM delivers the most consistent and significant performance gains, underscoring its efficacy in boosting personalized empathetic competencies.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC