AutoGen-Powered Multi-Agent Framework for Interpretable Multimodal Emotion Recognition

Emotion recognition allows intelligent systems to perceive and respond to human emotions. Single-modal methods often lack robustness, while multimodal approaches capture affective information more comprehensively. Existing multimodal models, however, struggle with modality heterogeneity, conflict resolution, and interpretability. To address this, we propose an AutoGen-based multi-agent framework, where expert agents for text, audio, video, fusion, and quality control collaborate through conflict-driven discussions. Experiments show it outperforms single-modal and conventional multimodal models in accuracy, F1-score, correlation, and robustness, while providing interpretable reasoning pathways.