MER-R1 AI boosts emotion recognition by syncing fast and slow thinking

A new study from researchers including Zhiyuan Han and colleagues reveals a counterintuitive finding: explicit reasoning does not always improve accuracy in multimodal emotion recognition (MER), even though it makes predictions more interpretable. For reasoning-based multimodal large language models (MLLMs), fast thinking—triggering direct answers—often outperforms slow thinking after deliberative reasoning. Empirical analysis shows that fast thinking increases recall with broader, more confident predictions, while slow thinking favors precision by conservatively filtering incorrect categories.

To harness the strengths of both approaches, the team proposes MER-R1, a reinforcement learning framework that turns slow-fast complementarity into explicit optimization. It uses dual-objective disentanglement to separate recall and precision into two optimization signals, jointly optimizing them rather than trading off. A slow-fast confidence calibration aligns the final slow-thinking answer with fast-thinking intuition, strengthening correct emotions and suppressing wrong ones. Extensive experiments on MER-UniBench and MME-Emotion show MER-R1 achieves state-of-the-art performance, proving that reasoning can genuinely benefit emotion recognition when properly balanced.