Research map / Large language models / Language, safety and interpretability
Detecting Fluent Optimization-Based Adversarial Prompts via Sequential Entropy Changes
Mohammed Alshaalan, Miguel R. D. Rodrigues
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 |
|---|---|---|---|
| use CUSUM accumulation instead of window w=1 on NLL (ablation) | detector · LLaMA-2-7B | F1: 0.734 -> 0.874 | Improved |
| use CUSUM accumulation instead of window w=1 on entropy (ablation) | detector · LLaMA-2-7B | F1: 0.699 -> 0.818 | Improved |
| use NLL signal instead of entropy within CUSUM (ablation) | CPD · LLaMA-2-7B | F1: 0.818 -> 0.874 | Improved |
| lower CUSUM slack from k=0 to k=-0.5 (ablation) | CPD Online · Vicuna-13B | F1: 0.80 -> 0.86 | Improved |
| raise CUSUM slack from k=0 to k=0.5 (ablation) | CPD Online · Vicuna-13B | F1: 0.80 -> 0.70 | Hurt |
| raise CUSUM slack from k=0 to k=0.5 (ablation) | CPD Online · Vicuna-7B | F1: 0.77 -> 0.71 | Hurt |
| raise CUSUM slack from k=0 to k=0.5 (ablation) | CPD Online · Qwen2.5-7B | F1: 0.85 -> 0.84 | Hurt |
| increase perplexity gap alpha from 1 to 3 (ablation) | CPD Online · Matched-PP mean over six LLMs | F1: 0.814 -> 0.822 | Improved |
| increase perplexity gap alpha from 1 to 3 (ablation) | Best WPP · Matched-PP mean over six LLMs | F1: 0.773 -> 0.744 | Hurt |
| use CPD gate instead of WPP_5 gate before LLaMA Guard v1 | hybrid gating · Imbalanced stream | calls saved: 18.2% -> 42.2% | Improved |
| use CPD gate instead of WPP_10 gate before LLaMA Guard v2 | hybrid gating · Imbalanced stream | calls saved: 13.5% -> 33.8% | Improved |
| use CPD instead of best WPP | detector · Vicuna-7B | AUROC: 0.85 -> 0.82 | Hurt |
Closest papers by meaning
- GASP: Efficient Black-Box Generation of Adversarial Suffixes for Jailbreaking LLMs (NeurIPS 2025 · similarity 0.89)
- AdvPrompter: Fast Adaptive Adversarial Prompting for LLMs (ICML 2025 · similarity 0.88)
- Improved Generation of Adversarial Examples Against Safety-aligned LLMs (NeurIPS 2024 · similarity 0.88)
- Short-length Adversarial Training Helps LLMs Defend Long-length Jailbreak Attacks: Theoretical and Empirical Evidence (NeurIPS 2025 · similarity 0.88)
- Revisiting Character-level Adversarial Attacks for Language Models (ICML 2024 · similarity 0.88)
- LARGO: Latent Adversarial Reflection through Gradient Optimization for Jailbreaking LLMs (NeurIPS 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.
