Beyond Model Collapse: Scaling Up with Synthesized Data Requires Reinforcement