Robust LLM safeguarding via refusal feature adversarial training

AAML
Re-assign community