SELF-EMO Framework Enables LLMs to Evolve Emotional Recognition and Expression Through Self-Play

A new research framework called SELF-EMO addresses limitations in emotional intelligence for large language models. The approach tackles both emotion recognition and coherent emotional expression, two capabilities essential for human-centric AI interactions. Current systems face constraints due to scarce, static annotated data. Grounded in the hypothesis that improved emotion prediction yields more consistent emotional responses, the framework introduces two auxiliary tasks: emotional understanding and emotional expression. A role-based self-play paradigm allows models to function as both emotion recognizers and dialogue responders. Through iterative interactions, diverse conversational trajectories are generated, enabling scalable data production. To maintain quality, a data flywheel mechanism filters candidate predictions and responses using a smoothed IoU-based reward. Selected high-quality outputs are then fed back into the training process. This self-evolution method aims to overcome data scarcity by creating its own training material. The work was announced on arXiv with the identifier 2604.18003v1. The research focuses on Emotion Recognition in Conversation (ERC) as a fundamental capability for advanced language models.