The capacity for empathy is becoming an increasingly crucial aspect of human-computer dialogue systems. Previous research has concentrated on the extraction of emotionally pertinent information from contextual data, However, there has been a lack of emphasis on assessing whether the responses truly align with the extracted information. This includes the emotional state of the speaker being addressed, the event being focused on, and the response strategies employed to deal with the event. To address this issue, we propose an innovative empathic response generation method with response attention guidance that leverages the instruction following capabilities of large language models. This allows the language model to focus more closely on the information that requires a response, and it has demonstrated superior capabilities in supporting emotions and empathy. The experiment results show that our method can generate coherent, informative and empathetic responses in both supervised fine-tuning and zero-shot settings, outperforming several baselines on two datasets.