AdvGRPO

New framework enhances adaptive red teaming for language models

Researchers have developed AdvGRPO, a novel co-training framework designed to enhance the adaptive red teaming of language models. This method addresses the instability of GRPO in attacker-defender optimization by emplo…