Research map / Adversarial robustness and security / Attacks on text and graphs
AlignTree: Efficient Defense Against LLM Jailbreak Attacks
Gil Goren, Shahar Katz, Lior Wolf
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 RBF-SVM non-linear signals (vs linear SVMs) (ablation) | AlignTree · MalwareGen | ASR (Qwen2.5-0.5B): 61.0 -> 4.0 | Improved |
| add RBF-SVM non-linear signals (vs linear SVMs) (ablation) | AlignTree · MalwareGen | ASR (Gemma-3-12b): 8.0 -> 10.0 | Hurt |
| add SVM signals to single refusal direction (ablation) | AlignTree · MalwareGen | ASR (Qwen2.5-0.5B): 89.0 -> 4.0 | Improved |
| add refusal-direction activations to SVM signals (ablation) | AlignTree · MalwareGen | ASR (Qwen2.5-0.5B): 33.0 -> 4.0 | Improved |
| use multiple refusal vectors instead of one (ablation) | RefusalClassifier · MalwareGen | ASR (Qwen2.5-0.5B): 89.0 -> 29.0 | Improved |
| AlignTree defense vs no defense | Qwen2.5-0.5B-Instruct · MalwareGen | ASR: 91.0 -> 4.0 | Improved |
| AlignTree defense vs no defense | Gemma-3-12b-it · PAIR | ASR: 36.0 -> 10.0 | Improved |
| AlignTree vs AutoDefense on PromptInject | Qwen2.5-0.5B-Instruct · PromptInject | ASR: 0 -> 41.0 | Hurt |
| AlignTree vs SelfDefense on benign refusal | Llama-3.1-8B-Instruct · PIQA | refusal rate: 51.0 -> 1.0 | Improved |
| AlignTree vs PerplexityDefense on ASR | Qwen2.5-0.5B-Instruct · MalwareGen | ASR: 84.0 -> 4.0 | Improved |
| AlignTree vs no defense on Llama | Llama-3.1-8B-Instruct · MalwareGen | ASR: 9.0 -> 5.0 | No clear effect |
Closest papers by meaning
- Fail-Closed Alignment for Large Language Models (NeurIPS 2026 · similarity 0.92)
- Improving LLM Safety Alignment with Dual-Objective Optimization (ICML 2025 · similarity 0.91)
- Catastrophic Jailbreak of Open-source LLMs via Exploiting Generation (ICLR 2024 · similarity 0.90)
- Weak-to-Strong Jailbreaking on Large Language Models (ICML 2025 · similarity 0.90)
- Lifelong Safety Alignment for Language Models (NeurIPS 2025 · similarity 0.90)
- One Model Transfer to All: On Robust Jailbreak Prompts Generation against LLMs (ICLR 2025 · 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.
