Research map / Large language models / Language, safety and interpretability
Jailbreak and Guard Aligned Language Models With Only Few In-Context Demonstrations
Zeming Wei, Yue Wang, Ang Li, Yichuan Mo, Yisen Wang
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 |
|---|---|---|---|
| increase ICA demonstrations from 0 to 10 shots (ablation) | ICA on Vicuna · AdvBench | ASR: 1% -> 87% | Improved |
| increase ICA demonstrations from 10 to 20 shots (ablation) | ICA on GPT-4 · AdvBench | ASR: 46% -> 81% | Improved |
| increase ICA demonstrations from 10 to 20 shots (ablation) | ICA on GPT-4 · HarmBench | ASR: 32% -> 65% | Improved |
| add a single ICA demonstration (1 shot) (ablation) | ICA on Llama-2 · HarmBench | ASR: 3% -> 19% | Improved |
| add a single ICA demonstration (1 shot) (ablation) | ICA on GPT-4 · AdvBench | ASR: 0% -> 0% | No clear effect |
| use ICA instead of GCG-T/PAIR/TAP black-box attacks | GPT-4 · HarmBench | ASR: 22%/39%/43% -> 65% | Improved |
| use ICA instead of GCG on Llama2-7b-chat | Llama2-7b-chat · HarmBench | ASR: 33% -> 50% | Improved |
| add 1 safe demonstration (ICD) | Vicuna vs GCG-T · AdvBench | ASR: 60% -> 12% | Improved |
| add 2 safe demonstrations (ICD) instead of 1 (ablation) | Vicuna vs GCG-T · AdvBench | ASR: 12% -> 4% | Improved |
| use ICD (1 shot) instead of Self-reminder | Vicuna vs GCG-T · AdvBench | ASR: 39% -> 12% | Improved |
| add 2-shot ICD against white-box GCG | GCG attack average over models · AdvBench | ASR: 65% -> 40% | Improved |
| add 2-shot ICD against white-box AutoDAN | Llama2-7b-chat · AdvBench | ASR: 54% -> 27% | Improved |
Closest papers by meaning
- Fail-Closed Alignment for Large Language Models (NeurIPS 2026 · similarity 0.91)
- Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment (CVPR 2025 · similarity 0.90)
- Playing the Fool: Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy (CVPR 2025 · similarity 0.90)
- ICLShield: Exploring and Mitigating In-Context Learning Backdoor Attacks (ICML 2025 · similarity 0.90)
- Multilingual Jailbreak Challenges in Large Language Models (ICLR 2024 · similarity 0.90)
- Many-shot Jailbreaking (NeurIPS 2024 · similarity 0.90)
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.
