Automated red teaming trains an attacker large language model (LLM) to elicit harmful responses from a victim LLM, but reinforcement learning can collapse at…
机构：Microsoft
来源：arXiv 2610.40926 | AI4Papers 论文推荐平台