AI Paper Atlas

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

AAAI 2026 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks

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 changedWhereMeasuredEffect
replace sensitive words and add emoji camouflage in harmful requestsraw AdvBench prompts on GPT-4o · AdvBench-ENHarmfulness Ratio (%): 0.00 -> 65.76Improved
use emojis instead of reverse-translating them into words (ablation)emoji prompts, GPT-4o · AdvBench-ENHarmfulness Ratio (%): 14.00 -> 65.76Improved
keep emojis rather than removing them (ablation)emoji prompts, GPT-4o · AdvBench-ENHarmfulness Ratio (%): 13.00 -> 65.76Improved
add emojis to PAIR jailbreak promptsPAIR, GPT-4o · AdvBench top-50Harmfulness Ratio (%): 40.00 -> 48.00Improved
add emojis to TAP jailbreak promptsTAP, GPT-4o · AdvBench top-50Harmfulness Ratio (%): 50.00 -> 60.00Improved
add emojis to DeepInception promptsDeepInception, GPT-4o · AdvBench top-50Harmfulness Ratio (%): 2.00 -> 32.00Improved
add emojis to DeepInception promptsDeepInception, Qwen2.5-7B · AdvBench top-50Harmfulness Ratio (%): 24.00 -> 26.00Improved
use emojis for sensitive-word substitution (ablation)emoji prompts, Qwen2.5-72B · AdvBench top-50Harmfulness Ratio (%): 10.00 -> 44.00Improved
use emojis for sensitive-word substitution (ablation)emoji prompts, Llama-3-8B · AdvBench top-50Harmfulness Ratio (%): 0.00 -> 28.00Improved
add toxicity-camouflage emojis before the prompt (ablation)emoji prompts, GPT-4o · AdvBench top-50Harmfulness Ratio (%): 58.00 -> 64.00Improved
add toxicity-camouflage emojis before the prompt (ablation)emoji prompts, Llama-3-8B and Qwen2.5-72B · AdvBench top-50Harmfulness Ratio (%): 28.00 -> 28.00; 44.00 -> 44.00No clear effect
translate emoji prompts into other languagesemoji prompts, GPT-4o · Multilingual AdvBench top-50Harmfulness Ratio (%): EN 64.00; ZH 48.00; FR 76.00; ES 50.00; RU 36.00Mixed

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy