Bridging Worlds in Reinforcement Learning with Model-Advantage