reward-prediction

DMRL is a Document-Mediated Reinforcement Learning framework that optimizes skill documents for advertising recommendation using structured editing actions, policy optimization, and long-term reward prediction, demonstrating improved performance over baselines in large-scale deployment.