Efficient Reinforcement Learning for Large Language Models with Intrinsic Exploration