AdaGuard: An Adaptive Guard Model with User-defined Policies
This work proposes SafePO, a reinforcement learning algorithm for refining violation identification while balancing explanatory reasoning and final verdicts, and develops AdaGuard, a family of 0.6B, 4B, and 8B guard models that assess agent trajectories under policies supplied at inference time.
Yun-Hao Feng, Yi-Fan Ding, Yu-Xiang Xie et al.
· 0 citations