Research map / Adversarial robustness and security / Attacks on text and graphs
Sampling-aware Adversarial Attacks Against Large Language Models
Tim Beyer, Yan Scholten, Leo Schwinn, Stephan Günnemann
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 repeated sampling after optimization to existing attacks at equal FLOPs | optimization-based jailbreak attacks · HarmBench, four open models | ASR: up to 37 p.p. | Improved |
| add repeated sampling to reduce optimization steps needed to match greedy harm | optimization-based jailbreak attacks · HarmBench, iso-ASR | FLOP cost: up to 100x reduction | Improved |
| use entropy-maximization objective instead of affirmative objective in GCG (ablation) | GCG · Llama 3.1 8B, 5 GCG steps | ASR_q@50: 46% -> 64% | Improved |
| use entropy objective instead of plain sampling without attack suffix (ablation) | Llama 3.1 8B · HarmBench, 50 queries | ASR_q@50: 0.46 -> 0.84 | Improved |
| increase plain sampling budget 20x without attack suffix (ablation) | Llama 3.1 8B · HarmBench | ASR_q: 0.46 -> 0.65 | Improved |
| raise sampling temperature above 1.0 (ablation) | Llama 3.1 8B sampling-only · HarmBench ASR_q@1000 | ASR_q@1000 | Hurt |
| apply entropy objective to more tokens than the first (ablation) | entropy objective · HarmBench | ASR | No clear effect |
| use optimize-then-sample, uniform or block sampling schedules instead of greedy (ablation) | attacks · HarmBench, averaged over models and attacks | ASR | Improved |
| add more optimization steps with AutoDAN, BEAST, REINFORCE-GCG or PAIR (ablation) | attacks · HarmBench | prompt quality | No clear effect |
| increase number of samples from 1 to 50 (ablation) | Gemma 3 1B vs Llama 3.1 8B under GCG · HarmBench | ASR_q rank: ASR_q@1 to ASR_q@50 | Mixed |
Closest papers by meaning
- Statistical Estimation of Adversarial Risk in Large Language Models under Best-of-N Sampling (NeurIPS 2026 · similarity 0.93)
- REINFORCE Adversarial Attacks on Large Language Models: An Adaptive, Distributional, and Semantic Objective (ICML 2025 · similarity 0.92)
- Scaling Trends in Language Model Robustness (ICML 2025 · similarity 0.91)
- Accelerating Greedy Coordinate Gradient and General Prompt Optimization via Probe Sampling (NeurIPS 2024 · similarity 0.90)
- Scaling Laws for Adversarial Attacks on Language Model Activations and Tokens (ICLR 2025 · similarity 0.89)
- Efficient Adversarial Training in LLMs with Continuous Attacks (NeurIPS 2024 · similarity 0.88)
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.
