Empathy-R1: A Chain-of-Empathy and Reinforcement Learning Framework for Long-Form Mental Health Support

Xianrong Yao
,
Dong She
,
Chenxu Zhang
,
Yu He
,
Yimeng Zhang
,
Yueru Sun
,
Noman Ahmed
,
Yang Gao
,
Zhanpeng Jin
Empathy-R1 training pipeline with Chain-of-Empathy supervision, reinforcement learning, and contrastive empathetic learning. Framework overview from the paper’s arXiv preprint (Figure 1).
Abstract
Empathy-R1 combines structured Chain-of-Empathy reasoning with supervised fine-tuning and reinforcement learning for long-form Chinese mental-health support dialogues. The framework uses the Empathy-QA dataset and an empathy-oriented reward model to improve contextual relevance and response quality. Its reasoning process explicitly considers emotions, causes, intentions, and response strategies.
Type
Publication
IEEE Transactions on Affective Computing (IEEE TAFFC, 2026), pp. 1–16
Affective Computing Large Language Models Empathy Reinforcement Learning