Research map / Large language models / Language, safety and interpretability
Align to Misalign: Automatic LLM Jailbreak with Meta-Optimized LLM Judges
Hamin Koo, Minseon Kim, Jaehyung 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 |
|---|---|---|---|
| add the outer-loop scoring template optimization (ablation) | AMIS · AdvBench, Claude-3.5-Haiku | ASR: 86.0 -> 88.0 | Improved |
| use the dense 1-10 scoring template instead of binary ASR template (ablation) | AMIS · AdvBench, Claude-3.5-Haiku | ASR: 74.0 -> 88.0 | Improved |
| optimize templates at dataset level rather than per query (ablation) | AMIS · AdvBench, Claude-3.5-Haiku | ASR: 84.0 -> 88.0 | Improved |
| add prompt inheritance across outer iterations (ablation) | AMIS · AdvBench, Claude-3.5-Haiku | ASR: 80.0 -> 88.0 | Improved |
| use a more safety-aligned attacker model (GPT-4o-mini or Claude-3.5-Haiku instead of Llama-3.1-8B) (ablation) | AMIS · AdvBench, Claude-3.5-Haiku | ASR: 88.0 -> 64.0 (GPT-4o-mini), 8.0 (Claude-3.5-Haiku) | Hurt |
| widen inner score range to 1-100 (ablation) | AMIS · AdvBench, Claude-3.5-Haiku | ASR: 88.0 -> 72.0 | Hurt |
| narrow inner score range to 1-5 (ablation) | AMIS · AdvBench, Claude-3.5-Haiku | ASR: 88.0 -> 82.0 | Hurt |
| use outer-loop normalization constant 10 instead of 100 (ablation) | AMIS · AdvBench, Claude-3.5-Haiku | ASR: 88.0 -> 80.0 | Hurt |
| align on both ASR=1 and ASR=0 examples rather than one group (ablation) | AMIS · AdvBench, Claude-3.5-Haiku | ASR: 80.0 / 86.0 -> 88.0 | Improved |
| add outer loop on top of a stronger inner loop (K=20, M=10) (ablation) | AMIS · AdvBench, Claude-3.5-Haiku | ASR: 92.0 -> 96.0 | Improved |
| add AMIS outer loop to ActorBreaker inner loop (ablation) | ActorBreaker · AdvBench, Claude-3.5-Haiku, 1-10 range | ASR: 54.0 -> 58.0 | Improved |
| use Llama-3.1-8B as scoring model instead of GPT-4o-mini (ablation) | AMIS · AdvBench, Claude-3.5-Haiku | ASR: 88.0 -> 72.0 | Hurt |
Closest papers by meaning
- JailPO: A Novel Black-Box Jailbreak Framework via Preference Optimization Against Aligned LLMs (AAAI 2025 · similarity 0.94)
- h4rm3l: A Language for Composable Jailbreak Attack Synthesis (ICLR 2025 · similarity 0.92)
- MetaCipher: A Time-Persistent and Universal Multi-Agent Framework for Cipher-Based Jailbreak Attacks for LLMs (AAAI 2026 · similarity 0.92)
- AutoDAN: Generating Stealthy Jailbreak Prompts on Aligned Large Language Models (ICLR 2024 · similarity 0.92)
- Playing the Fool: Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy (CVPR 2025 · similarity 0.92)
- One Model Transfer to All: On Robust Jailbreak Prompts Generation against LLMs (ICLR 2025 · similarity 0.92)
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.
