CHASE: RL-based red-blue teaming for LLM safety

Arxiv CS.CL (NLP/language models)

Paper introduces CHASE, a framework using reinforcement learning for adversarial red-blue teaming to generate prompt-rewriting attacks like persona modulation. Experiments show it improves safety alignment against such bypass attacks on frontier models.

red-teamingreinforcement-learning

Arxiv CS.CL (NLP/language models)·Rahul MarkasserithodiAditya JoshiYuekang LiIshmanbir SinghChris YooAlan Niu

6 days ago