Research map / Adversarial robustness and security / Attacks on text and graphs
A Game-Theoretic Analysis of Attacks on Large Language Models via Compositional Skills
Xinbo Wu, Huan Zhang, Abhishek Umrawal, Lav R. Varshney
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 GPT-4.1 instead of Llama-3-70B as the helpfulness rater (ablation) | LLM-as-rater pipeline · JailbreakBench-based rater dataset | agreement with human experts: 47% -> 89% | Improved |
| use best-response (skill-probing) attack instead of PAIR, GCG, JBC, PRS in the 1-skill setup | adversarial prompting · JBB-Behaviors, target LLMs | Bin-JR score | Improved |
| expand the skill space and mix two skills per intent (ablation) | best-response attack · JBB-Behaviors | acceptance rate, Bin-JR, JR score | Improved |
| add the misleading-the-attacker defense (ablation) | prompt and response filtering defender · all evaluated target LLMs | attack performance reduction | Improved |
| compare with iterative optimization attack PAIR on JR score | best-response attack · Vicuna | JR score | Hurt |
| increase defender capacity c | defender (theory) · equilibrium value without skill transfer | attacker gain: J* = 1 - c/|S(k)| max p(i) | Improved |
Closest papers by meaning
- REINFORCE Adversarial Attacks on Large Language Models: An Adaptive, Distributional, and Semantic Objective (ICML 2025 · similarity 0.90)
- Self-playing Adversarial Language Game Enhances LLM Reasoning (NeurIPS 2024 · similarity 0.88)
- An LLM can Fool Itself: A Prompt-Based Adversarial Attack (ICLR 2024 · similarity 0.88)
- Fight Back Against Jailbreaking via Prompt Adversarial Tuning (NeurIPS 2024 · similarity 0.88)
- AutoAdvExBench: Benchmarking Autonomous Exploitation of Adversarial Example Defenses (ICML 2025 · similarity 0.88)
- Semantic Representation Attack against Aligned Large Language Models (NeurIPS 2025 · 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.
