Confidence Aware Reinforcement Learning: Advancing Large Language Models in Dynamic Environments