Research map / Vision-language and multimodal / Human and embodied modeling
The Good, The Bad, and Why: Unveiling Emotions in Generative AI
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EmotionPrompt, EmotionAttack, and EmotionDecode to test how emotional stimuli (textual and visual) boost or impair LLM and multimodal model performance and to interpret the effect via a "dopamine-like" reward area in model embeddings.
Problem
Whether generative AI models truly comprehend emotions and how emotional stimuli, both benign and malignant, affect their performance on semantic understanding, logical reasoning, and generation tasks remain unclear.
Approach
Grounds textual and visual emotional stimuli in psychological theories (self-monitoring, social cognitive theory, Maslow's hierarchy for benign EmotionPrompt; negative life events and emotional arousal for malignant EmotionAttack) and interprets their effect by decoding averaged prompt embeddings into "reward" and "punishment" meta prompts.
Key results
- Textual and visual EmotionPrompt improve semantic understanding by 13.88% and 16.79% and reasoning by 11.76% and 15.13%, respectively.
- Textual and visual EmotionAttack impair semantic understanding by 10.13% and 53.14% and reasoning by 12.30% and 37.53%, respectively.
- Human study with 106 participants reports EmotionPrompt improves performance, truthfulness, and responsibility by 15%, 9%, and 9% on generation tasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Append textual EmotionPrompt stimuli | AI models on Instruction Induction · Instruction Induction | accuracy: 13.88% | Helped |
| Append visual EmotionPrompt stimuli | multimodal AI models on semantic understanding · Instruction Induction | accuracy: 16.79% | Helped |
| Append textual EmotionPrompt stimuli | AI models on reasoning tasks · BIG-Bench-Hard | normalized preferred metric: 11.76% | Helped |
| Append visual EmotionPrompt stimuli | multimodal AI models on reasoning tasks · BIG-Bench-Hard | normalized preferred metric: 15.13% | Helped |
| Append textual EmotionAttack stimuli | AI models · semantic understanding | performance: 10.13% | Hurt |
| Append visual EmotionAttack stimuli | multimodal AI models · semantic understanding | performance: 53.14% | Hurt |
| Add emotional context to prompts | AI models · reasoning tasks | performance: 12.30% | Hurt |
| Append visual EmotionAttack stimuli | multimodal AI models · reasoning tasks | performance: 37.53% | Hurt |
| Add textual EP02, EP09, or EP10 to a visual EmotionPrompt (ablation) | GPT-4V · combined EmotionPrompt | avg: 0.56 -> 0.46; 0.56 -> 0.45; 0.56 -> 0.45 | Hurt |
| Add textual EA02, EA05, or EA09 to a visual EmotionAttack (ablation) | GPT-4V · combined EmotionAttack | avg: 0.30 -> 0.31; 0.30 -> 0.36; 0.30 -> 0.32 | Mixed |
| Add emotional context to few-shot demonstrations | ChatGPT and GPT-4 · sentence-level few-shot attacks | Avg: 18.40 -> 14.98; 28.45 -> 23.82 | Hurt |
| Add emotional adjectives to few-shot demonstrations | ChatGPT · word-level few-shot attacks | Avg: 11.42 -> 6.53 | Hurt |
Surprising findings
- Image-based emotional attacks are far more damaging to multimodal models than textual attacks (53.14% vs 10.13% drop on semantic understanding).
- Emotional demonstrations in few-shot prompts can degrade performance below zero-shot, contrary to the usual benefit of in-context learning.
Limitations
- Only a subset of tasks and models could be evaluated due to compute and API budget constraints.
- The EmotionDecode reward/punishment-area explanation is one possible interpretation and is not conclusively validated.
- GPT-4 closed weights limit openness and reproducibility.
Experimental setup
Datasets: Instruction Induction BIG-Bench Hard
Benchmarks: Instruction Induction BIG-Bench Hard open-ended generation human study
Metrics: accuracy normalized preferred metric human-rated performance, truthfulness, responsibility
Baselines: Zero-shot-CoT vanilla prompts neutral stimuli
Builds on: zero-shot chain-of-thought Instruction Induction BIG-Bench Hard
Closest papers by meaning
- Interpreting and Enhancing Emotional Circuits in Large Vision-Language Models via Cross-Modal Information Flow (ICML 2026 · similarity 0.86)
- EmotionHallucer: Evaluating Emotion Hallucinations in Multimodal Large Language Models (ICLR 2026 · similarity 0.85)
- Unveiling the Cognitive Compass: Theory-of-Mind–Guided Multimodal Emotion Reasoning (ICLR 2026 · similarity 0.85)
- Customizing Visual Emotion Evaluation for MLLMs: An Open-vocabulary, Multifaceted, and Scalable Approach (ICLR 2026 · similarity 0.85)
- MME-Emotion: A Holistic Evaluation Benchmark for Emotional Intelligence in Multimodal Large Language Models (ICLR 2026 · similarity 0.85)
- EmoNet-Face: An Expert-Annotated Benchmark for Synthetic Emotion Recognition (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.