Evaluating the Validity of Word-level Adversarial Attacks with Large Language Models F i n d i n g s - A C L