MAGIC: A Co-Evolving Attacker-Defender Adversarial Game for Robust LLM Safety
Ziyu Wan, Chaochao Lu, Ying Wen et al. · 2 Feb 2026
A novel multi-turn multi-agent reinforcement learning framework formulates LLM safety alignment as an adversarial asymmetric game, enabling co-evolution of attacker and defender agents to discover and mitigate unseen attack patterns.