We apply an information-theoretic perspective to reconsider generative document retrieval (GDR), in which a document x∈X is indexed by t∈T, and a neural autoregressive model is trained to map queries Q to T. GDR…