A learning signal is information derived from observed outcomes that guides model learning or parameter updates. In retrieval-augmented generation, reader performance can supervise a retriever even when no document-level relevance labels exist. section 6.2.1
A reader's probability of the correct answer supplies the observable outcome. Retrieved documents remain latent variables: documents that make the answer easier to generate, reflected by lower perplexity, receive stronger retriever scores. This feedback loop addresses irrelevant retrieval that a static retriever cannot correct. section 6.2.1
Retrieval feedback without relevance labels
A retriever supplies several documents to a reader. If one document helps the reader generate the correct answer, the training objective increases that document's retrieval score, without requiring a human-annotated query-document pair. section 6.2.1



