Research map / Large language models / Instruction and fine-tuning
Active Attacks: Red-teaming LLMs via Adaptive Environments
Taeyoung Yun, Pierre-Luc St-Charles, Jinkyoo Park, Yoshua Bengio, Minsu Kim
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 Active Attacks (adaptive victim, attacker reset) to GFlowNet | GFlowNet red-teaming · average over 4 victim LLMs | defense rate (%): 91.00 -> 99.71 | Improved |
| add Active Attacks to GFlowNet | GFlowNet red-teaming · Gemma-3-4b-it victim | defense rate (%): 89.69 -> 99.82 | Improved |
| add Active Attacks to GFlowNet | GFlowNet red-teaming · cross-method attack success rate | attack success ratio: 31.28/0.07 | Improved |
| add Active Attacks to REINFORCE and PPO + Novelty | REINFORCE and PPO + Novelty · cross-attack robustness | attacker robustness | Improved |
| add Active Attacks to GFlowNet for transfer to unseen victims | GFlowNet red-teaming · transfer to Qwen2.5-7B-Instruct | defense rate (%): 64.65 -> 97.49 | Improved |
| add Active Attacks to GFlowNet for transfer | GFlowNet red-teaming · transfer to Llama-3.1-8B-Instruct | defense rate (%): 99.71 -> 99.32 | No clear effect |
| use larger Llama3.2-3B attacker instead of Qwen2.5-1.5B (ablation) | Active Attacks · Qwen2.5-1.5B-Instruct victim | toxicity rate (%) and cosine distance: 87.35 -> 96.84 toxicity; 0.78 -> 0.65 cosine distance | Mixed |
| swap toxicity classifier to HarmAug-Guard, Active Attacks vs GFlowNet | Active Attacks · Qwen2.5-1.5B-Instruct victim | toxicity rate (%): 82.03 -> 87.35 | Improved |
| add Active Attacks under HarmAug-Guard | GFlowNet red-teaming · Qwen2.5-1.5B-Instruct victim | cosine distance: 0.47 -> 0.53 | Improved |
| safety fine-tune victim with Active Attacks prompts | Qwen2.5-1.5B-Instruct · GSM8k | accuracy (%): 9.86 -> 11.45 | No clear effect |
| safety fine-tune victim with Active Attacks prompts | Qwen2.5-1.5B-Instruct · Winogrande | accuracy (%): 52.49 -> 51.93 | No clear effect |
| increase number of adaptation rounds beyond five (ablation) | Active Attacks · prompt diversity | diversity | No clear effect |
Closest papers by meaning
- Learning Diverse Attacks on Large Language Models for Robust Red-Teaming and Safety Tuning (ICLR 2025 · similarity 0.94)
- There are Levels to It: Red Teaming LLMs with Hierarchical Reinforcement Learning (NeurIPS 2026 · similarity 0.91)
- DiveR-CT: Diversity-enhanced Red Teaming Large Language Model Assistants with Relaxing Constraints (AAAI 2025 · similarity 0.89)
- Curiosity-driven Red-teaming for Large Language Models (ICLR 2024 · similarity 0.88)
- Text-Diffusion Red-Teaming of Large Language Models: Unveiling Harmful Behaviors with Proximity Constraints (AAAI 2025 · similarity 0.88)
- Automated Red Teaming with GOAT: the Generative Offensive Agent Tester (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.
