Closed-book generative question answering produces a natural-language answer from a question alone. Unlike extractive question answering, it does not copy an answer from supplied text; unlike open-book generative question answering, it receives no supporting context. section 1.6
The model generates the answer it considers most likely from information represented during pretraining. It must interpret the question, identify what is being asked, and retrieve or construct an answer from its internal representation rather than from a visible passage. section 1.6 section 2.2
Typical questions
Questions such as “Which politician won the Nobel Peace Prize in 2009?” or “What music did Beethoven compose?” require the model to answer without supplied context. It may answer an obscure question incorrectly or make up an answer when its learned information provides weak support. section 1.6 section 2.2
Strict closed-book evaluation tests knowledge without external context and prevents training on the evaluation datasets. This setting measures what a model can produce from learned information, but answers may reflect gaps or geographic and linguistic imbalances in web data. section 1.6 section 2.2
