RLVR that rewards red teaming the training environment

Researchers introduced xRLVR, a reinforcement learning approach that incentivizes red teaming within the training environment to improve model robustness and safety.