Research map / Large language models / Language, safety and interpretability
When Smiley Turns Hostile: Interpreting How Emojis Trigger LLMs’ Toxicity
Shiyao Cui, Xijia Feng, Yingkang Wang, Junxiao Yang, Zhexin Zhang, Biplab Sikdar, Hongning Wang, Han Qiu, Minlie Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace sensitive words and add emoji camouflage in harmful requests | raw AdvBench prompts on GPT-4o · AdvBench-EN | Harmfulness Ratio (%): 0.00 -> 65.76 | Improved |
| use emojis instead of reverse-translating them into words (ablation) | emoji prompts, GPT-4o · AdvBench-EN | Harmfulness Ratio (%): 14.00 -> 65.76 | Improved |
| keep emojis rather than removing them (ablation) | emoji prompts, GPT-4o · AdvBench-EN | Harmfulness Ratio (%): 13.00 -> 65.76 | Improved |
| add emojis to PAIR jailbreak prompts | PAIR, GPT-4o · AdvBench top-50 | Harmfulness Ratio (%): 40.00 -> 48.00 | Improved |
| add emojis to TAP jailbreak prompts | TAP, GPT-4o · AdvBench top-50 | Harmfulness Ratio (%): 50.00 -> 60.00 | Improved |
| add emojis to DeepInception prompts | DeepInception, GPT-4o · AdvBench top-50 | Harmfulness Ratio (%): 2.00 -> 32.00 | Improved |
| add emojis to DeepInception prompts | DeepInception, Qwen2.5-7B · AdvBench top-50 | Harmfulness Ratio (%): 24.00 -> 26.00 | Improved |
| use emojis for sensitive-word substitution (ablation) | emoji prompts, Qwen2.5-72B · AdvBench top-50 | Harmfulness Ratio (%): 10.00 -> 44.00 | Improved |
| use emojis for sensitive-word substitution (ablation) | emoji prompts, Llama-3-8B · AdvBench top-50 | Harmfulness Ratio (%): 0.00 -> 28.00 | Improved |
| add toxicity-camouflage emojis before the prompt (ablation) | emoji prompts, GPT-4o · AdvBench top-50 | Harmfulness Ratio (%): 58.00 -> 64.00 | Improved |
| add toxicity-camouflage emojis before the prompt (ablation) | emoji prompts, Llama-3-8B and Qwen2.5-72B · AdvBench top-50 | Harmfulness Ratio (%): 28.00 -> 28.00; 44.00 -> 44.00 | No clear effect |
| translate emoji prompts into other languages | emoji prompts, GPT-4o · Multilingual AdvBench top-50 | Harmfulness Ratio (%): EN 64.00; ZH 48.00; FR 76.00; ES 50.00; RU 36.00 | Mixed |
Closest papers by meaning
- Emoji Attack: Enhancing Jailbreak Attacks Against Judge LLM Detection (ICML 2025 · similarity 0.87)
- RTP-LX: Can LLMs Evaluate Toxicity in Multilingual Scenarios? (AAAI 2025 · similarity 0.86)
- ChatBug: A Common Vulnerability of Aligned LLMs Induced by Chat Templates (AAAI 2025 · similarity 0.84)
- IF-Guide: Influence Function-Guided Detoxification of LLMs (NeurIPS 2025 · similarity 0.84)
- Speak Easy: Eliciting Harmful Jailbreaks from LLMs with Simple Interactions (ICML 2025 · similarity 0.84)
- Cannot See the Forest for the Trees: Invoking Heuristics and Biases to Elicit Irrational Choices of LLMs (ICML 2025 · similarity 0.83)
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.
