AI Paper Atlas

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

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Safety alignment after fine-tuning

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 changedWhereMeasuredEffect
add Active Attacks (adaptive victim, attacker reset) to GFlowNetGFlowNet red-teaming · average over 4 victim LLMsdefense rate (%): 91.00 -> 99.71Improved
add Active Attacks to GFlowNetGFlowNet red-teaming · Gemma-3-4b-it victimdefense rate (%): 89.69 -> 99.82Improved
add Active Attacks to GFlowNetGFlowNet red-teaming · cross-method attack success rateattack success ratio: 31.28/0.07Improved
add Active Attacks to REINFORCE and PPO + NoveltyREINFORCE and PPO + Novelty · cross-attack robustnessattacker robustnessImproved
add Active Attacks to GFlowNet for transfer to unseen victimsGFlowNet red-teaming · transfer to Qwen2.5-7B-Instructdefense rate (%): 64.65 -> 97.49Improved
add Active Attacks to GFlowNet for transferGFlowNet red-teaming · transfer to Llama-3.1-8B-Instructdefense rate (%): 99.71 -> 99.32No clear effect
use larger Llama3.2-3B attacker instead of Qwen2.5-1.5B (ablation)Active Attacks · Qwen2.5-1.5B-Instruct victimtoxicity rate (%) and cosine distance: 87.35 -> 96.84 toxicity; 0.78 -> 0.65 cosine distanceMixed
swap toxicity classifier to HarmAug-Guard, Active Attacks vs GFlowNetActive Attacks · Qwen2.5-1.5B-Instruct victimtoxicity rate (%): 82.03 -> 87.35Improved
add Active Attacks under HarmAug-GuardGFlowNet red-teaming · Qwen2.5-1.5B-Instruct victimcosine distance: 0.47 -> 0.53Improved
safety fine-tune victim with Active Attacks promptsQwen2.5-1.5B-Instruct · GSM8kaccuracy (%): 9.86 -> 11.45No clear effect
safety fine-tune victim with Active Attacks promptsQwen2.5-1.5B-Instruct · Winograndeaccuracy (%): 52.49 -> 51.93No clear effect
increase number of adaptation rounds beyond five (ablation)Active Attacks · prompt diversitydiversityNo clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy