Enhancing the Capability and Robustness of Large Language Models through Reinforcement Learning-Driven Query Refinement