Affective-CoT: Decomposing Multimodal Emotion Reasoning through a Hierarchical Cognitive Workflow

摘要

Multimodal Emotion Recognition (MER) has advanced significantly with the advent of Multimodal Large Language Models (MLLMs), which enable generative, descriptive understanding of complex human affect. However, the inherent ''black-box'' nature of these end-to-end models limits their trustworthiness and applicability in high-stakes domains, particularly due to their opacity in handling conflicting cross-modal cues (e.g., sarcasm). To address this critical gap, we propose Affective-CoT, a novel hierarchical framework that explicitly decouples perception from reasoning to achieve interpretable and faithful emotion analysis. Our framework utilizes specialized perception models to extract structured semantic evidence from raw audiovisual streams, which is then integrated and arbitrated by a central reasoning LLM executing a meticulously designed Cognitive Workflow. Critically, Affective-CoT generates a nuanced emotion description grounded in a transparent, human-interpretable reasoning trace. The efficacy of our framework was decisively validated by securing first place in the official MER-2025 Descriptive Emotion Understanding (DES) challenge. This result not only highlights the superiority of our method but also champions a new paradigm for building scrutable and trustworthy affective computing systems.