UniESC: A Personality-Sensitive Multi-Agent Role-Playing for Emotional Support Conversation

Emotional Support Conversation (ESC) refers to dialogue systems designed to provide empathetic responses and psychological aid to users facing emotional distress. However, due to dataset limitations, existing efforts in ESC often produce generic responses that fail to address users' personalized needs. To overcome this challenge, we propose UniESC, a personality sensitive multi-agent role-playing framework for delivering tailored emotional support. At the core of our framework is a Theory-of-Mind (ToM)-based Roundtable mechanism, in which School Supporters, agents guided by ToM, collaboratively simulate discussions customized to the seeker's characteristics. A supervisor then selects the most appropriate response to form the UniConv dataset. Notably, UniESC leverages existing ESC datasets instead of relying solely on synthetic data. We further fine-tune a compact model, UniChat, using UniConv to provide personalized support. Experimental results show that UniChat achieves the best overall human evaluation score of 2.87% and strong automatic performance, including METEOR score of 26.05%, ROUGE-L score of 22.28%, and Vector Extrema score of 96.03%, demonstrating the effectiveness of our framework.