DARE to Disagree: A Multi-Agent Adversarial Debate Framework for Open-Vocabulary Multimodal Emotion Recognition

Open-vocabulary multimodal emotion recognition (OV-MER) is fundamentally challenged by the inherent ambiguity of human expression. Monolithic Multimodal Large Language Models (MLLMs), while powerful, struggle to reconcile the generative freedom required for this task with the need for high semantic precision, often leading to outputs that are incoherent or logically inconsistent. To resolve this conflict, we propose DARE (Debating Agents for Robust Emotion-understanding), a novel framework that reframes OV-MER as a structured, multi-agent adversarial debate. Our implementation, the Emotion Adversarial Debate Framework (EAD-Frame), orchestrates a multi-turn dialectic between a 'liberal' and a 'conservative'agent, moderated by a 'judge' to synthesize a robust, fine-grained consensus from conflicting initial hypotheses. On the highly competitive MER2025 Challenge, EAD-Frame, implemented with an open-source MLLM, outperforming single-agent baselines including top-tier models like GPT-4o. Our unsupervised proxy metrics further confirm that DARE's outputs are significantly more coherent and exhibit a drastically lower contradiction rate. We present a new, interpretable paradigm for managing ambiguity, offering a robust and generalizable solution for complex reasoning tasks across the AI landscape.