Tag
AdaGuard introduces a family of guard models (0.6B, 4B, 8B) that assess language model agent trajectories under user-defined policies, using a new dataset and reinforcement learning algorithm for adaptive safety.