MER-R1: Multimodal Emotion Reasoning via Slow-Fast Thinking Synergy

Short summary

MER-R1 unifies fast thinking (high recall) and slow reasoning (high precision) for multimodal emotion recognition via reinforcement learning. The framework uses dual-objective optimization and confidence calibration to achieve state-of-the-art results on emotion benchmarks, proving that fast-slow synergy outperforms either approach alone.

• •Fast vs. slow thinking trade-off in emotion recognition: fast favors recall, slow favors precision
• •MER-R1 framework uses dual-objective disentanglement to jointly optimize both signals
• •Achieves SOTA on MER-UniBench and MME-Emotion benchmarks with theoretical justification

Generated with AI, which can make mistakes.