LLMs-as-Instructors: Learning from Errors Toward Automating Model Improvement F i n d i n g s - E M N L P