LLM-Empowered State Representation for Reinforcement Learning