SELF-EMO: Emotional Self-Evolution from Recognition to Consistent Expression

Emotion Recognition in Conversation (ERC) has become a fundamental capability for large language models (LLMs) in human-centric interaction. Beyond accurate recognition, recent studies highlight the importance of coherent emotional expression, yet both abilities are severely limited by the scarcity and static nature of high-quality annotated data.In this work, we propose SELF-EMO, a self-evolution framework grounded in a psychologically motivated hypothesis: the better a model predicts others’ emotions, the better it can generate its own emotionally consistent responses. Building on this insight, we explicitly incorporate two auxiliary tasks—emotional understanding and emotional expression—and formulate a role-based self-play paradigm in which the model simultaneously acts as an emotion recognizer and a dialogue responder. Through iterative interaction, the agents continuously generate diverse conversational trajectories, forming a scalable data generation process. To ensure training quality, we construct a data flywheel mechanism: each self-play rollout produces multiple candidate emotional predictions and responses, which are then filtered via a smoothed IOU-based reward to select high-quality samples. The selected samples are fed back into training, enabling continuous self-improvement without external supervision. Based on this paradigm, we further develop SELF-GRPO, a reinforcement learning algorithm that stabilizes optimization under diverse emotional outputs by combining multi-label alignment rewards with group-level consistency signals.Extensive experiments on three benchmark datasets (IEMOCAP, MELD, and EmoryNLP) demonstrate that SELF-EMO achieves state-of-the-art (SOTA) performance without relying on external retrieval or auxiliary models. Under a unified training setting, our method improves average accuracy by +6.33% on Qwen3-4B and +8.54% on Qwen3-8B, with consistent gains across all benchmarks. These results validate the effectiveness of the proposed self-evolution paradigm and highlight its strong generalization ability.