Skip to content

Paper 011 / DeepMind

Improving language models by retrieving from trillions of tokens

RETRO augments a language model with nearest-neighbor chunks retrieved from a database containing trillions of tokens, reducing reliance on parameters alone.

UPDATED 16 Jul 2026SOURCE-LED REVIEWRESEARCH REVIEW
01

Paper, researchers, and primary source

Major lab research / retrieval_augmented_generation

RETRO augments a language model with nearest-neighbor chunks retrieved from a database containing trillions of tokens, reducing reliance on parameters alone.

CONTRIBUTION / 01

Contribution 1

Combined chunk-level retrieval with cross-attention inside a language model.

CONTRIBUTION / 02

Contribution 2

Scaled retrieval over a two-trillion-token database.

CONTRIBUTION / 03

Contribution 3

Showed retrieval-augmented models can match much larger dense models on studied tasks.

02

Research context

retrieval_augmented_generation / 2021

Improving language models by retrieving from trillions of tokens places retro inside the broader retrieval augmented generation discussion at DeepMind, with retrieval supplying a second analytical lens. The paper's through-line contains three reported moves: Combined chunk-level retrieval with cross-attention inside a language model; Scaled retrieval over a two-trillion-token database; and Showed retrieval-augmented models can match much larger dense models on studied tasks. That sequence keeps external memory tied to the reported work instead of treating it as a free-standing promise. The practical stake is equally bounded: external memory can improve factual coverage and updateability, but retrieval quality and source governance become part of model quality.

03

Methods and evidence reading

1 cataloged method notes

METHOD / 01

Method 1

The experimental design in Improving language models by retrieving from trillions of tokens tests combined chunk-level retrieval with cross-attention inside a language model and scaled retrieval over a two-trillion-token database against the paper's documented baselines, datasets, model variants, or systems workloads.

How to read the evidence

Evidence for Improving language models by retrieving from trillions of tokens is best read as a scoped argument rather than a universal verdict. The source record contains two linked moves: Combined chunk-level retrieval with cross-attention inside a language model; and Scaled retrieval over a two-trillion-token database. Its reported outcome is: Showed retrieval-augmented models can match much larger dense models on studied tasks. The defensible takeaway remains external memory can improve factual coverage and updateability, but retrieval quality and source governance become part of model quality. That conclusion must travel with the recorded boundary that evidence for retro in Improving language models by retrieving from trillions of tokens covers evaluation protocol, comparison baselines, documented data, compute budget, architecture choices, and task distribution; behavior beyond that documented envelope remains untested. A replication should preserve the disclosed setup and test whether retro still holds when retrieval conditions change. Testing Improving language models by retrieving from trillions of tokens beyond its source setting requires a stable retro control, explicit retrieval slices, and documented exceptions to Scaled retrieval over a two-trillion-token database.

04

Findings in the source record

1 paper-specific findings

  1. The reported evidence in Improving language models by retrieving from trillions of tokens supports showed retrieval-augmented models can match much larger dense models on studied tasks.
05

Practical implication for AI builders

DeepMind / 2021

06

Proposed BrokenGPT application

Research blueprint / proposed status

INTEGRATION POINT / 01

Proposed route placement / retro

For a proposed BrokenGPT experiment based on Improving language models by retrieving from trillions of tokens, add an optional retrieval tier whose responses expose retrieved passages, provenance, and retrieval latency separately from generation latency. Keep the retro path isolated, versioned, and attributable to this research record.

VALIDATION METRIC / 02

Proposed acceptance test / retrieval

Validate the proposed retro route against the paper's reported outcome: Showed retrieval-augmented models can match much larger dense models on studied tasks. Measure attribution, answer support, retrieval recall, and latency for the Improving language models by retrieving from trillions of tokens candidate, then isolate retrieval regressions before judging the proposed retro route.

TRADEOFF / 03

Proposed decision boundary / external memory

Balance fresh evidence, index cost, and retrieval failure before promoting the proposed external memory design. Because even if the reported result reproduces, a later model revision, another user population, changed operating conditions, a different product, and new hardware can reverse its product value and must be measured separately, adoption remains conditional on replay under BrokenGPT's selected model, runtime, and policy configuration.

07

Limitations, verification, and source

Boundaries recorded with the paper

Limitations

  • Evidence for retro in Improving language models by retrieving from trillions of tokens covers evaluation protocol, comparison baselines, documented data, compute budget, architecture choices, and task distribution; behavior beyond that documented envelope remains untested.
  • Even if the reported result reproduces, a later model revision, another user population, changed operating conditions, a different product, and new hardware can reverse its product value and must be measured separately.

PRIMARY SOURCES

  1. 01
    Improving language models by retrieving from trillions of tokens

    DeepMind — Primary primary arXiv paper / 8 December 2021 / Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, and 25 more

Related research reviews

View all 100 credited research papers

STRAIGHT ANSWERS

Frequently asked questions

01What does Improving language models by retrieving from trillions of tokens study?

RETRO augments a language model with nearest-neighbor chunks retrieved from a database containing trillions of tokens, reducing reliance on parameters alone.

02Which methods does Improving language models by retrieving from trillions of tokens use?

The experimental design in Improving language models by retrieving from trillions of tokens tests combined chunk-level retrieval with cross-attention inside a language model and scaled retrieval over a two-trillion-token database against the paper's documented baselines, datasets, model variants, or systems workloads.

03What does Improving language models by retrieving from trillions of tokens report?

The reported evidence in Improving language models by retrieving from trillions of tokens supports showed retrieval-augmented models can match much larger dense models on studied tasks.

04What is the proposed BrokenGPT application for Improving language models by retrieving from trillions of tokens?

Proposed: add an optional retrieval tier whose responses expose retrieved passages, provenance, and retrieval latency separately from generation latency.

MAJOR LAB RESEARCH / PAPER 011

Continue after Improving language models by retrieving from trillions of tokens

After Improving language models by retrieving from trillions of tokens, browse the full index for adjacent retrieval_augmented_generation research and work from DeepMind.

Open the paper index