The rise of large language models (LLMs) has significantly advanced open-domain dialogue systems, yet their ability to convey empathy remains limited, particularly in emotionally sensitive contexts such as counseling or support services. This raises important questions: Can LLMs engage in emotional interactions with users as naturally and meaningfully as a human conversation? Can they truly understand and respond with empathy, making users feel genuinely heard and supported? To address these challenges, we introduce EmpLLM, a framework designed to enhance LLMs’ empathetic capabilities through a combination of psychologist role-play and inner contemplation. We developed a high-quality dialogue dataset for model training and proposed a new evaluation benchmark, EmpTest, to assess the model’s empathy and emotional intelligence. Experimental results demonstrate that EmpLLM significantly improves the model’s ability to engage in emotionally responsive conversations, offering a promising path toward more human-like and empathetic conversational agents.