DACE: Diversity-Driven Adversarial Co-Evolution for Robust LLM Safety Alignment
DACE: Diversity-Driven Adversarial Co-Evolution for Robust LLM Safety Alignment
AI summary
DACE is a diversity-driven adversarial co-evolution framework that targets attack strategy collapse and defensive adversarial forgetting to enhance the robustness of LLM safety alignment. On the attack side, it uses a two-dimensional strategy space of risk category × attack style together with a marginal coverage gain reward; on the defense side, it uses Thompson sampling to maintain a Bayesian adversarial replay pool. On standard safety, automated attacker, and general capability benchmarks, DACE improves robustness to out-of-distribution attacks, preserves general capabilities, and achieves broader attack strategy coverage.