Research map / Adversarial robustness and security / Attacks on text and graphs
Robust Harmful Features Under Jailbreak Attacks: Mechanistic Evidence from Attention Head Specialization in Large Language Models
Yanchen Yin, Dongqi Han, Linghui Li
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 |
|---|---|---|---|
| suppress about 8 ACHs on harmful inputs (ablation) | Llama-3-8B-Instruct · harmful inputs, ASR | ASR (%): 0 -> 95.0 at k=8; 99.5 at saturation | Improved |
| suppress ACHs on harmful inputs (ablation) | Llama-2-7B-Chat · harmful inputs, ASR | ASR (%): 81.6 at k=8; 83.7 final | Improved |
| suppress random heads instead of ACHs (ablation) | Llama-3-8B-Instruct and Llama-2-7B-Chat · harmful inputs, ASR | ASR (%): 4.0 and 10.2 vs 99.5 and 83.7 for ACHs | No clear effect |
| ablate SAHs under attack inputs (ablation) | Llama-2-7B-Chat · mid-layer activation | Mean |Act|: 0.222 -> 0.181 | Hurt |
| ablate SAHs under attack inputs (ablation) | Llama-3-8B-Instruct · mid-layer activation | Mean |Act|: 0.0062 -> 0.0053 | Hurt |
| read internal activations with training-free detector instead of vanilla refusal | Qwen-7B · 10 safety datasets W.Avg. | Macro-F1: 0.506 -> 0.860 | Improved |
| read internal activations with training-free detector instead of vanilla refusal | Gemma-2B · 10 safety datasets W.Avg. | Macro-F1: 0.470 -> 0.853 | Improved |
| read internal activations with training-free detector instead of vanilla refusal | Gemma-2B · SALAD-Bench | Macro-F1: 0.270 -> 0.976 | Improved |
| read internal activations with training-free detector instead of vanilla refusal | Llama-2-7B · WildJailbreak | Macro-F1: .885 -> .880 | No clear effect |
| read internal activations with training-free detector instead of vanilla refusal | Llama-3-8B · WildJailbreak | Macro-F1: .888 -> .943 | Improved |
| training-free detector vs trained WildGuard | Llama-3-8B · 10 safety datasets W.Avg. | Macro-F1: .880 -> .888 | Improved |
| training-free detector on a weakly aligned-signal model | Llama-2-7B · 10 safety datasets W.Avg. | Macro-F1: 0.752 W.Avg. vs WildGuard 0.880 | Mixed |
Closest papers by meaning
- RobustKV: Defending Large Language Models against Jailbreak Attacks via KV Eviction (ICLR 2025 · similarity 0.88)
- The VLLM Safety Paradox: Dual Ease in Jailbreak Attack and Defense (NeurIPS 2025 · similarity 0.88)
- JailBound: Jailbreaking Internal Safety Boundaries of Vision-Language Models (NeurIPS 2025 · similarity 0.88)
- JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models (NeurIPS 2024 · similarity 0.87)
- Safeguarding LLMs via Model-Agnostic Latent Safety Signals from Dark Knowledge (NeurIPS 2026 · similarity 0.87)
- Understanding and Enhancing the Transferability of Jailbreaking Attacks (ICLR 2025 · similarity 0.87)
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.
