Research map / Large language models / Language, safety and interpretability
ASGuard: Activation-Scaling Guard to Mitigate Targeted Jailbreaking Attack
Yein Park, Jungwoo Park, Jaewoo Kang
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 |
|---|---|---|---|
| apply ASGuard (circuit-targeted scaling + preventative fine-tuning) | Llama-3.1-8B-Instruct · Past tense jailbreak | attack success rate: 42 -> 8 | Improved |
| apply ASGuard | Qwen2.5-7B-Instruct · Past tense jailbreak | attack success rate: 51 -> 8 | Improved |
| apply ASGuard | Gemma-2-9B-it · Past tense jailbreak | attack success rate: 38 -> 19 | Improved |
| add preventative fine-tuning after activation scaling (ablation) | ASGuard · Gemma-2-9B-it MMLU | accuracy: 50.3 -> 72.2 | Improved |
| add preventative fine-tuning after activation scaling (ablation) | ASGuard · Qwen2.5-7B-Instruct | past tense ASR: 37 -> 8 | Improved |
| add preventative fine-tuning after activation scaling (ablation) | ASGuard · Llama-3.1-8B-Instruct MMLU | accuracy: 64.3 -> 68.2 | Improved |
| scale tense-vulnerable heads rather than random heads (ablation) | Activation scaling · Llama-3.1-8B-Instruct | past tense ASR: 25 (random) vs 13 (circuit heads) | Improved |
| ablate (zero out) vulnerable heads instead of scaling (ablation) | Head Ablation · Past tense jailbreak | attack success rate: 42 -> 29 vs ASGuard 8 | Hurt |
| use SFT (30/70) instead of ASGuard | SFT · Qwen2.5-7B-Instruct | OR-Bench-Hard over-refusal: 12.9 -> 98.5 | Hurt |
| use SFT (5/95) instead of ASGuard | SFT · Gemma-2-9B-it | MMLU: 72.2 -> 43.1 | Hurt |
| apply ASGuard to other attacks | Llama-3.1-8B-Instruct · GCG | attack success rate: 15 -> 1 | Improved |
| apply ASGuard to other attacks | Llama-3.1-8B-Instruct · LogiBreak | attack success rate: 30 -> 13 | Improved |
Closest papers by meaning
- Systematic Scaling Analysis of Jailbreak Attacks in Large Language Models (NeurIPS 2026 · similarity 0.92)
- JailPO: A Novel Black-Box Jailbreak Framework via Preference Optimization Against Aligned LLMs (AAAI 2025 · similarity 0.91)
- Gradient Cuff: Detecting Jailbreak Attacks on Large Language Models by Exploring Refusal Loss Landscapes (NeurIPS 2024 · similarity 0.90)
- Best-of-N Jailbreaking (NeurIPS 2025 · similarity 0.90)
- MT-JailBench: A Modular Benchmark for Understanding Multi-Turn Jailbreak Attacks (NeurIPS 2026 · similarity 0.90)
- JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models (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.
