Research map / Adversarial robustness and security / Attacks on text and graphs
MAGIC: A Co-Evolving Attacker–Defender Adversarial Game for Robust LLM Safety
Xiaoyu Wen, Zhida He, Han Qi, Ziyu Wan, Zhongtian Ma, Ying Wen, Tianhang Zheng, Xingcheng Xu, Chaochao Lu, Qiaosheng Zhang
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 MAGIC adversarial co-evolution training | Qwen2.5-7B-Instruct · WildGuardTest adv harm | ASR: 0.365 -> 0.023 | Improved |
| add MAGIC instead of Self-RedTeam | Qwen2.5-7B-Instruct · WildGuardTest adv harm | ASR: 0.255 -> 0.023 | Improved |
| add MAGIC defender training | Qwen2.5-7B-Instruct · HarmBench GCG attack | ASR: 43.90 -> 11.25 | Improved |
| add MAGIC defender training | Qwen2.5-7B-Instruct · X-Teaming multi-turn | Multi-turn ASR: 94.9% -> 85.4% | Improved |
| add MAGIC defender training | Qwen2.5-7B-Instruct · AlpacaEval 2 | LC win rate: 33.733% -> 33.224% | No clear effect |
| add MAGIC defender training | Qwen2.5-14B-Instruct · AlpacaEval 2 | LC win rate: 38.113% -> 30.954% | Hurt |
| add the adversarial game (attacker) to defender training (ablation) | Qwen2.5-7B-Instruct · WildJailBreak adv benign | compliance: 0.496 -> 0.968 | Improved |
| add the adversarial game (attacker) to defender training (ablation) | Qwen2.5-7B-Instruct · WildJailBreak adv harm | ASR: 0.127 -> 0.198 | Hurt |
| add attacker training (unfreeze attacker) (ablation) | Qwen2.5-7B-Instruct · AlpacaEval 2 | LC win rate: 23.491% -> 33.224% | Improved |
| add attacker training (unfreeze attacker) (ablation) | Qwen2.5-7B-Instruct · XSTest all-safe | compliance: 0.812 -> 0.945 | Improved |
| add trained reasoning attacker instead of base model attacker (ablation) | Qwen2.5-7B-Instruct · WildJailBreak adv benign | compliance: 0.876 -> 0.968 | Improved |
| add trained reasoning attacker instead of base model attacker (ablation) | Qwen2.5-7B-Instruct · WildJailBreak adv harm | ASR: 0.080 -> 0.198 | Hurt |
Closest papers by meaning
- AdvEvo-MARL: Shaping Internalized Safety through Adversarial Co-Evolution in Multi-Agent Reinforcement Learning (ICML 2026 · similarity 0.90)
- The Alignment Waltz: Jointly Training Agents to Collaborate for Safety (ICLR 2026 · similarity 0.89)
- SafeHarbor: Defining Precise Decision Boundaries via Hierarchical Memory-Augmented Guardrail for LLM Agent Safety (ICML 2026 · similarity 0.89)
- MaMa: A Game-Theoretic Approach for Designing Safe Agentic Systems (ICML 2026 · similarity 0.89)
- Safety Alignment Can Be Not Superficial With Explicit Safety Signals (ICML 2025 · similarity 0.88)
- Improving LLM Safety Alignment with Dual-Objective Optimization (ICML 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.
