Tag
DRACO is a reinforcement learning method that dynamically generates rubrics and redistributes trajectory scores to improve long-horizon agent performance without verifiers.