Automating Dataset Updates Towards Reliable and Timely Evaluation of Large Language Models