Reproduce “The Good, The Bad, and Why: Unveiling Emotions in Generative AI” 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
Cheng Li, Jindong Wang, Yixuan Zhang, Kaijie Zhu, Xinyi Wang, Wenxin Hou, Jianxun Lian, Fang Luo and 2 more
Researchers from Microsoft Research and collaborating institutions explored how generative AI models comprehend and are influenced by emotional stimuli, proposing EmotionPrompt to enhance performance, EmotionAttack to impair it, and EmotionDecode for interpretability. The work demonstrated that textual EmotionPrompt improved semantic understanding by 13.88% and logical reasoning by 11.76%, while visual EmotionAttack decreased semantic understanding by 53.14%.
Key insights
- Generative AI models consistently understand and are significantly influenced by both benign and malignant emotional stimuli.
- Multi-modal models are more sensitive to emotional stimuli, especially visual inputs, suggesting greater vulnerability to visual EmotionAttack.
- AI models exhibit computational 'reward' and 'punishment' areas, particularly in deeper layers, where emotional stimuli induce effects analogous to dopamine mechanisms in humans, influencing attention weights.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.