Skip to content

Learning to Coordinate: LLM-Guided Workflow Evolution for Token-Efficient Red Teaming

Learning to Coordinate: LLM-Guided Workflow Evolution for Token-Efficient Red Teaming

arXiv: 其他 AI 安全论文· arXiv:2611.24221· Xiaoyang Xing, Wenyu Dai, Xiaoliang Gong, Asoke K Nandi, Anthony Cohn·Indexed by this site · 13 hr ago· Originally publishedAttacks· Its place in the scan record →
AI summary

L2C uses LLM-guided multi-agent workflow evolution to discover LLM safety failures at lower token cost. A supervised fine-tuned controller decides, based on the target response and evaluation verdict, whether to keep, refine, or restructure the workflow. Across six configurations of three executors and two target models on HarmBench, L2C raises attack success rate by 1.4 to 4.5 percentage points over EvoSynth, with average online tokens reduced by 35.6% to 37.3%.

Read originalarxiv.org