paper / lilianweng / May 13 / failedDiverse and Effective Red Teaming with Auto-generated Rewards and Multi-step Reinforcement Learning