본문으로 건너뛰기
Safety Alignment as Continual Learning: Mitigating the Alignment Tax via Orthogonal Gradient Projection | AI Trends