The paper proposes a framework called "coverage–compression," which abstracts the closed-book question-answering task as a setting of N queries, K answers per query, M facts observed, compressed into at most B bits.