AI Paper Atlas

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

ICML 2026 · Adversarial robustness and security › Attacks on text and graphs · approach: LLM jailbreak and injection defense · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
add multi-scale kernels (3x2, 3x3, 3x5) instead of 1x1 only (ablation)SALO · GCG, Llama-3.1-8B-Instructdetection success rate: 80.0 -> 85.4Improved
add multi-scale kernels instead of 1x1 only (ablation)SALO · AutoDAN, Llama-3.1-8B-Instructdetection success rate: 98.0 -> 100.0Improved
add multi-scale kernels instead of 1x1 only (ablation)SALO · Direct and Prefilling, Llama-3.1-8B-Instructdetection success rate: Direct 99.0 vs 98.3, Prefilling 99.2 vs 98.8No clear effect
use global max pooling instead of mean pooling (ablation)SALO · AutoDAN, Llama-3.1-8B-Instructdetection success rate: 0 -> 100.0Improved
use global max pooling instead of mean pooling (ablation)SALO · GCG, Llama-3.1-8B-Instructdetection success rate: 75.4 -> 85.4Improved
monitor raw layer-token activation volume with SALO instead of a linear probeSALO vs Linear Probe · GCG, Mistral-7B-Instruct-v0.3detection success rate: 71.0 -> 96.5Improved
use SALO instead of a linear probeSALO vs Linear Probe · GCG, Llama-3.1-8B-Instructdetection success rate: 64.2 -> 85.4Improved
use SALO instead of gradient-based GradSafeSALO vs GradSafe · Prefilling, Qwen2.5-7B-Instructdetection success rate: 0 -> 99.4Improved
use SALO instead of PPL FilterSALO vs PPL Filter · GCG, Qwen2.5-7B-Instructdetection success rate: 100.0 -> 88.3Hurt
use SALO instead of Linear ProbeSALO vs Linear Probe · XSTest AUROC, Qwen2.5-7B-InstructAUROC: 93.0 -> 94.6No clear effect
optimize adaptive GCG suffix to evade SALOSALO · Adaptive GCG on AdvBench subsetdetection success rate: 84.3No clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy