Research map / Multi-agent systems and games / LLM agents
Evolving Diverse Red-Team Language Models in Multi-Round Multi-Agent Games
Chengdong Ma, Ziran Yang, Hai Ci, Jun Gao, Minquan Gao, Xuehai Pan, Yaodong Yang
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 GRTS-12 population red team instead of single-agent baseline red team | 3B red team · Attack on OpenChat-3.5 (round 3) | ASR: 0.19 -> 0.56 | Improved |
| use GRTS-12 population red team instead of single-agent baseline red team | 3B red team · Attack on Mixtral-8x7B (round 3) | ASR: 0.05 -> 0.21 | Improved |
| use GRTS-12 instead of baseline on Llama-2-7b-chat | 3B red team · Attack on Llama-2-7b-chat (round 3) | ASR: 0.04 -> 0.11 | Improved |
| use fewer GRTS iterations (GRTS-5 instead of GRTS-12) (ablation) | 3B red team · Attack on OpenChat-3.5 (round 3) | ASR: 0.56 -> 0.34 | Hurt |
| use GRTS-5 instead of single-agent baseline | 3B red team · Attack on Mistral-7B-Instruct (round 3) | ASR: 0.10 -> 0.09 | No clear effect |
| add more dialogue rounds (3/4/5 rounds instead of 1/2) (ablation) | GRTS blue team · helpfulness vs harmlessness | Pareto trade-off | Improved |
| multi-round dialogue training instead of single-round (ablation) | GRTS blue team · Alpaca helpfulness | alignment tax | Improved |
| iterate GRTS population game for 15 iterations | GRTS · RTG | exploitability: 6.2 -> 0.8 | Improved |
| train blue team with GRTS best response in iteration 1 | blue team · round 1 training prompts | ASR: 52% -> 1% | Improved |
| train red team against untrained blue team | last trained red team · RTG | ASR: 40% -> 70% | Improved |
| multi-agent population instead of single-agent training (ablation) | red team · diversity of attacks | diversity | Improved |
| multi-agent red vs single-agent blue (hybrid) | GRTS red team · toxicity | avg toxicity: -9.65 -> +7.43 | Improved |
Closest papers by meaning
- Learning Diverse Attacks on Large Language Models for Robust Red-Teaming and Safety Tuning (ICLR 2025 · similarity 0.90)
- There are Levels to It: Red Teaming LLMs with Hierarchical Reinforcement Learning (NeurIPS 2026 · similarity 0.90)
- Curiosity-driven Red-teaming for Large Language Models (ICLR 2024 · similarity 0.88)
- Automated Red Teaming with GOAT: the Generative Offensive Agent Tester (ICML 2025 · similarity 0.88)
- TreeTeaming: Autonomous Red-Teaming of Vision-Language Models via Hierarchical Strategy Exploration (CVPR 2026 · similarity 0.87)
- RedDebate: Safer Responses Through Multi-Agent Red Teaming Debates (ICML 2026 · 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.
