Thinking about how to extract: Energizing LLMs’ emergence capabilities for document-level event argument extraction F i n d i n g s - A C L