Knowledge

ClashEval

ClashEval: quantifying the tug-of-war between an LLM's internal prior and external (retrieved) evidence. Each QA item is answered twice — closed-book (prior) and with a perturbed reference passage (post) whose gold answer has been mutated. Responses are binary correctness against the true answer across six domains (drug dosages, Olympic records, names, locations, dates, news).