Training AI models to attack and defend themselves simultaneously

Original: Learning to Attack and Defend: Adaptive Red Teaming of Language Models via GRPO

Source: arxiv.org ↗