Adaptive Instruction Composition Red-Teaming: RL-Guided Attack Diversity Beats Random Combination

arXiv·medium signal

Capital One's AI Foundations team combines crowdsourced harmful queries and tactics using RL to jointly optimize attack effectiveness and diversity, rather than random combination. The RL mechanism balances exploration with exploitation in the combinatorial instruction space. Substantially outperforms random combination on effectiveness and diversity metrics, even under model transfer to unseen targets.