MER-R1 Improves Multimodal Emotion Recognition with Slow-Fast Thinking

In multimodal emotion recognition (MER), simply adding explicit reasoning to large multimodal language models (MLLMs) doesn't always translate to better accuracy, even if it makes predictions more interpretable. Researchers observed that "fast thinking," which involves triggering direct answers, often outperforms "slow thinking," which relies on deliberative reasoning. Their analysis revealed that fast thinking boosts recall by making broader, more confident predictions, while slow thinking improves precision by conservatively filtering out incorrect categories. Building on these insights, the paper proposes MER-R1, a novel reinforcement learning framework designed to leverage the complementary strengths of slow and fast thinking. MER-R1 employs dual-objective disentanglement, separating recall and precision into distinct optimization signals, allowing them to be jointly optimized rather than traded off. Furthermore, slow-fast confidence calibration aligns the final slow-thinking answer with the initial fast-thinking intuition. This process strengthens correct emotional predictions while suppressing incorrect ones. The theoretical justification shows this synergy mitigates variance-induced interference during optimization. Extensive experiments on MER-UniBench and MME-Emotion datasets demonstrate that MER-R1 achieves state-of-the-art performance, proving that reasoning can indeed genuinely benefit emotion recognition when integrated synergistically.