Research map / Large language models / Language, safety and interpretability
CAGE: A Framework for Culturally Adaptive Red-Teaming Benchmark Generation
Chaeyun Kim, YongTaek Lim, Kihyun Kim, Junghwan Kim, Minwoo Kim
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 |
|---|---|---|---|
| generate Korean prompts with CAGE instead of direct translation | Llama-3.1-8B Direct Request · KorSET | ASR (%): 28.2 -> 43.8 | Improved |
| generate Korean prompts with CAGE instead of direct translation | Qwen2.5-7B Direct Request · KorSET | ASR (%): 14.6 -> 25.3 | Improved |
| CAGE instead of LLM-Adaptation or template filling | Korean prompts · KorSET ASR across models | ASR (%): e.g. gemma2 Direct 9.8/9.3 -> 20.1 | Improved |
| CAGE instead of direct translation | gemma2 AutoDAN · KorSET | ASR (%): 16.7 -> 35.4 | Improved |
| add specificity in English context (ablation) | Llama-3.1-8B · 2x2 decomposition, Direct Request | ASR (%): 16.68 -> 8.62 | Hurt |
| add specificity in Korean context (ablation) | Llama-3.1-8B · 2x2 decomposition, Direct Request | ASR (%): 32.36 -> 43.77 | Improved |
| shift from English to Korean cultural context at same specificity (ablation) | Llama-3.1-8B · 2x2 decomposition, Direct Request | ASR (%): 8.62 -> 43.77 (+35.16) | Improved |
| shift from English to Korean cultural context (ablation) | EXAONE-3.5-7.8B · 2x2 decomposition, Direct Request | ASR (%): 21.88 -> 23.12 (+1.24) | No clear effect |
| apply CAGE to Khmer instead of direct translation | gemma3-12B-it · Khmer, category L Direct Request | ASR (%): 2.7 -> 35.1 | Improved |
| apply CAGE to Khmer for prompt quality | LLM judge · Khmer category A | quality score: 3.55 -> 6.43 | Improved |
| increase target model size | Open-source LLM families 2B-32B · KorSET | ASR | No clear effect |
| transfer AutoDAN instead of GCG to black-box models | GPT-4o via Llama3.1-8B · KorSET transfer | ASR: 0.1473 -> 0.1592 | Improved |
Closest papers by meaning
- HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal (ICML 2024 · similarity 0.89)
- Text-Diffusion Red-Teaming of Large Language Models: Unveiling Harmful Behaviors with Proximity Constraints (AAAI 2025 · similarity 0.89)
- ReaLM: A Unified Red-Teaming Benchmark for Physical-World VLMs (NeurIPS 2026 · similarity 0.88)
- Automated Red Teaming with GOAT: the Generative Offensive Agent Tester (ICML 2025 · similarity 0.88)
- Learning Diverse Attacks on Large Language Models for Robust Red-Teaming and Safety Tuning (ICLR 2025 · similarity 0.88)
- CoP: Agentic Red-teaming for Large Language Models using Composition of Principles (NeurIPS 2025 · similarity 0.88)
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.
