Borgeaud and colleagues: RETRO chunked retrieval-enhanced language modeling

What changed

Conditioned generation on retrieved text chunks using a frozen retriever and trainable encoding and cross-attention, testing large external text stores.

Scientific scope

Conditioned generation on retrieved text chunks using a frozen retriever and trainable encoding and cross-attention, testing large external text stores. Retrieval-augmented generation is inherited; the contribution is the chunked architecture and scale study. Database storage, retrieval costs and data overlap remain relevant.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Retrieval-augmented generation is inherited; the contribution is the chunked architecture and scale study. Database storage, retrieval costs and data overlap remain relevant.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Improving language models by retrieving from trillions of tokens

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction