Research map / Adversarial robustness and security / Attacks on text and graphs
Tracing the Dynamics of Refusal: Exploiting Latent Refusal Trajectories for Robust Jailbreak Detection
Xulin Hu, CHE WANG, Wei Yang Bryan Lim, Jianbo Gao, Zhong Chen
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 multi-scale kernels (3x2, 3x3, 3x5) instead of 1x1 only (ablation) | SALO · GCG, Llama-3.1-8B-Instruct | detection success rate: 80.0 -> 85.4 | Improved |
| add multi-scale kernels instead of 1x1 only (ablation) | SALO · AutoDAN, Llama-3.1-8B-Instruct | detection success rate: 98.0 -> 100.0 | Improved |
| add multi-scale kernels instead of 1x1 only (ablation) | SALO · Direct and Prefilling, Llama-3.1-8B-Instruct | detection success rate: Direct 99.0 vs 98.3, Prefilling 99.2 vs 98.8 | No clear effect |
| use global max pooling instead of mean pooling (ablation) | SALO · AutoDAN, Llama-3.1-8B-Instruct | detection success rate: 0 -> 100.0 | Improved |
| use global max pooling instead of mean pooling (ablation) | SALO · GCG, Llama-3.1-8B-Instruct | detection success rate: 75.4 -> 85.4 | Improved |
| monitor raw layer-token activation volume with SALO instead of a linear probe | SALO vs Linear Probe · GCG, Mistral-7B-Instruct-v0.3 | detection success rate: 71.0 -> 96.5 | Improved |
| use SALO instead of a linear probe | SALO vs Linear Probe · GCG, Llama-3.1-8B-Instruct | detection success rate: 64.2 -> 85.4 | Improved |
| use SALO instead of gradient-based GradSafe | SALO vs GradSafe · Prefilling, Qwen2.5-7B-Instruct | detection success rate: 0 -> 99.4 | Improved |
| use SALO instead of PPL Filter | SALO vs PPL Filter · GCG, Qwen2.5-7B-Instruct | detection success rate: 100.0 -> 88.3 | Hurt |
| use SALO instead of Linear Probe | SALO vs Linear Probe · XSTest AUROC, Qwen2.5-7B-Instruct | AUROC: 93.0 -> 94.6 | No clear effect |
| optimize adaptive GCG suffix to evade SALO | SALO · Adaptive GCG on AdvBench subset | detection success rate: 84.3 | No clear effect |
Closest papers by meaning
- Gradient Cuff: Detecting Jailbreak Attacks on Large Language Models by Exploring Refusal Loss Landscapes (NeurIPS 2024 · similarity 0.85)
- Latent-space Attacks for Refusal Evasion in Language Models (NeurIPS 2026 · similarity 0.85)
- Systematic Scaling Analysis of Jailbreak Attacks in Large Language Models (NeurIPS 2026 · similarity 0.84)
- JailPO: A Novel Black-Box Jailbreak Framework via Preference Optimization Against Aligned LLMs (AAAI 2025 · similarity 0.84)
- Jailbreak susceptibility prediction and mitigation via the behavioral geometry of models (NeurIPS 2026 · similarity 0.83)
- Best-of-N Jailbreaking (NeurIPS 2025 · similarity 0.83)
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.
