Reproduce “MER-R1: Multimodal Emotion Reasoning via Slow-Fast Thinking Synergy” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhiyuan Han, Beier Zhu, Wenwen Tong, Chengwei Qin, Xinyi Wang, Jiayu Zhang, Jiangnan Chen, Hewei Guo and 3 more

The MER-R1 framework integrates "slow thinking" and "fast thinking" modes in Multimodal Large Language Models to improve emotion recognition accuracy, addressing a common paradox where explicit reasoning often underperforms intuitive responses. The model achieved a new state-of-the-art mean score of 83.50 on the MER-UniBench and 51.5 on MME-Emotion benchmarks, demonstrating that deliberative reasoning can genuinely enhance predictive capability.

Key insights

  • "Fast thinking" exhibits broader emotion coverage and higher recall with strong confidence in correct categories, while "slow thinking" provides more concentrated, precise predictions and effectively suppresses incorrect categories.
  • The "thinking paradox" stems from standard F1 optimization methods in RL that can bias towards objectives with larger normalized variation, leading to detrimental trade-offs between recall and precision for slow thinking.
  • Explicitly disentangling recall and precision objectives, combined with calibrating slow-thinking confidence by leveraging fast-thinking's intuitive strengths, can resolve this paradox and make deliberative reasoning genuinely beneficial for accuracy.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.