Optimizing Long-Context Processing with Role-RL: A Reinforcement Learning Framework for Efficient Large Language Model Deployment