Learning Object States from Actions via Large Language Models