MER-R1: Multimodal Emotion Reasoning via Slow-Fast Thinking Synergy
06:00 · June 29, 2026 · arXiv cs.AI RSS
We find that explicit reasoning does not necessarily translate into better multimodal emotion recognition (MER) accuracy, even though it makes predictions more interpretable. Specifically, for reasoning-based MLLMs, fast thinking by triggering direct answers often outperforms slow thinking after deliberative reasoning. Our empirical analyses show that fast thinking improves recall with broader and more confident predictions, whereas slow thinking favors precision through conservative filtering of incorrect categories. Building on these insights, we propose MER-R1, a reinforcement learning framework that turns slow-fast complementarity into explicit optimization. Dual-objective disentanglement separates recall and precision into two optimization signals, allowing them to be jointly optimized rather than traded off against each other. Slow-fast confidence calibration further aligns the final slow-thinking answer with fast-thinking intuition, strengthening correct emotions while suppressing incorrect ones. In this way, MER-R1 unifies the recall-oriented intuition of fast thinking with the precision-oriented selectivity of slow thinking. We further provide theoretical justification for this synergy, showing that it mitigates variance-induced interference during optimization. Extensive experiments on MER-UniBench and MME-Emotion show that MER-R1 achieves state-of-the-art performance and makes reasoning genuinely benefit emotion recognition.
Summary
The article examines a counter-intuitive finding in multimodal emotion recognition: explicit chain-of-thought reasoning in multimodal large language models often fails to raise accuracy even though it improves interpretability. On open-vocabulary benchmarks, direct “fast thinking” answers without reasoning consistently achieve higher recall than the more deliberative “slow thinking” traces that current reinforcement-learning-with-verifiable-reward methods produce. Fast thinking yields broader, more confident predictions across emotion categories, whereas slow thinking applies conservative filtering that raises precision at the cost of suppressing some correct labels.
MER-R1 converts this observed complementarity into an explicit optimization target. Its dual-objective disentanglement maintains separate reward and advantage signals for recall and precision, preventing the variance-induced bias that arises when a single F1 objective is normalized across groups. A subsequent slow-fast confidence calibration step then transfers the higher on correct categories from fast thinking to the final slow-thinking output, while preserving the suppression of incorrect categories already achieved by slow reasoning. Theoretical analysis shows that the resulting joint optimization reduces interference between the two objectives during policy updates.
Experiments on MER-UniBench and MME-Emotion confirm that the framework reaches state-of-the-art recognition performance while ensuring that the generated reasoning chains genuinely contribute to accuracy rather than serving only as post-hoc explanations.
Why it matters
This research is highly relevant for Dutch AI researchers focusing on multimodal LLMs, affective computing, and interpretable AI. The exploration of explicit reasoning mechanisms aligns with the Netherlands' focus on transparent AI, though the application of emotion recognition requires careful consideration under the EU AI Act.
