Papers › Improving language models by retrieving from trillions of tokens

Improving language models by retrieving from trillions of tokens

8 Dec 2021arXiv:2112.04426archive 2025-07-28

Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George van den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, Diego de Las Casas, Aurelia Guy, Jacob Menick, Roman Ring, Tom Hennigan, Saffron Huang, Loren Maggiore, Chris Jones, Albin Cassirer, Andy Brock, Michela Paganini, Geoffrey Irving, Oriol Vinyals, Simon Osindero, Karen Simonyan, Jack W. Rae, Erich Elsen, Laurent SIfre

We enhance auto-regressive language models by conditioning on document chunks retrieved from a large corpus, based on local similarity with preceding tokens. With a 2 trillion token database, our Retrieval-Enhanced Transformer (RETRO) obtains comparable performance to GPT-3 and Jurassic-1 on the Pile, despite using 25× fewer parameters. After fine-tuning, RETRO performance translates to downstream knowledge-intensive tasks such as question answering. RETRO combines a frozen Bert retriever, a differentiable encoder and a chunked cross-attention mechanism to predict tokens based on an order of magnitude more data than what is typically consumed during training. We typically train RETRO from scratch, yet can also rapidly RETROfit pre-trained transformers with retrieval and still achieve good performance. Our work opens up new avenues for improving language models through explicit memory at unprecedented scale.

PaperPDFCodeCode Syntology ranMethod as code

In Syntology View this paper on Syntology, its page in Syntology's graph. That page adds which of its functions are also present in other papers' code Syntology holds, with dates, and its dated run record as JSON and a README badge. It also lists the repositories linked to the paper, the abstract and the calls for agents.

Open this paper in Syntology's Atlas, the map of the papers in Syntology's graph and their citations.

For agents, on Syntology's MCP service (how to connect):

Code

Syntology 23 code samples: 20 from 2 repositories linked to this paper, 3 identical to code harvested from another paper's repository. Across all 23: 16 ran (of which 5 constructed an object rather than computing a result; 14 with no instrument failure: 0 honoured, 3 violated, 11 with no contract checked; 2 where Syntology's instrument failed) · 7 unverified. Where Syntology's instrument failed: 2 where our contract draft was wrong; failures of Syntology's instrument, not of the code.

By repository: community (archive-listed): 20 samples from 2 repositories, 13 ran (11 with no instrument failure; 2 where Syntology's instrument failed); 3 identical to code first harvested elsewhere. The run record, sample by sample. “Ran” means executed on a synthesized input, not that the code is correct or reproduces the paper.

Levels written on 2026-09-22 (UTC): the dates Syntology's graph wrote each sample's verification level, not when the code executed. 3 of the 16 samples that ran have no recorded date.

lucidrains/RETRO-pytorch mentioned on GitHubpytorchSyntology: 13 samples, 8 ran (2 instrument-failed) · 5 unverified report
labmlai/annotated_deep_learning_paper_implementations pytorchSyntology: 7 samples, 5 ran (0 instrument-failed) · 2 unverified report

Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.

Code Syntology ran Syntology

23 samples harvested: 16 ran (of which 5 constructed an object rather than computing a result; 14 with no instrument failure: 0 honoured, 3 violated, 11 with no contract checked; 2 where Syntology's instrument failed) · 7 unverified. “Where Syntology's instrument failed” counts failures of Syntology's instrument, not of the code. Syntology's record for this page has not changed since , the first build that kept a record date for it; when this build read Syntology's graph is in the build record. That date is not when the samples ran.

3ran · violated contract
2ran · our draft was wrong
11ran
7unverified

Licence: 3 of the 23 samples are pointer only, meaning Syntology does not serve that copy's text. This page shows no code text for any sample; each one links to its file in the repository.

Harvested from 2 repositories linked to this paper, official or community; the list below is grouped by repository, each group headed by its name and what it is to the paper. 3 samples are listed last, in a group without a repository. “Ran” means the sample executed on a synthesized input. It does not mean the output is correct, and nothing here reproduces the paper's results. A line beginning “fixture:” under a sample that ran gives the arguments Syntology's fixture called the sample with, each as name:type/rank/element kind (? where none was recorded), and the type of what the fixture's run returned, exactly as the harness recorded them; the arguments and the returned type were recorded by two separate runs of the same fixture. No output value is shown, and a returned type says nothing about whether the output was right. “Level recorded” gives the UTC date when Syntology's graph wrote the sample's verification level, not when the code executed; where the level was written before Syntology's graph kept that date, the row says “level recorded, date not recorded”. For 3 samples that ran, the graph records a signature this page does not show: for a class the fixture built without calling, the harness called the object with inputs it made up itself, so the recorded signature is the harness's probe, not the fixture's arguments. “Honoured” and “violated” refer to a contract Syntology drafted from the code itself. “Our draft was wrong” and “the contract check could not run” are failures of Syntology's instrument, not of the code: the sample ran on Syntology's synthesized input, and then the drafted contract did not describe it, or the check itself raised an error (RAISES), lacked a dependency (DEP_MISSING) or ran out of time (TIMEOUT); hover the label for which. A metamorphic tier, where shown, is evidence of the output's shape, not that the sample implements the method.

Each sample ends with its code_sha256. A sample's code_sha256 is Syntology's identity for that exact code. It is printed in full (16 hexadecimal characters for every sample in this build), and get_code(code_sha256="…") takes it exactly as printed. An agent calls it on Syntology's MCP service (how to connect); click an identity to copy that call.

Repository labels, per sample. official repository: The archive marks this repository official for the paper. named in the paper: The archive records that the paper mentions this repository; it is not marked official. community (archive-listed): In the archive's code links for this paper, not marked official and not recorded as mentioned in the paper. found in paper text by Syntology: Syntology found this repository in the paper's own text; whether it is the authors' implementation is not asserted. community: Not in the archive's code links for this paper; a community repository Syntology harvested. Samples are grouped by repository: a repository marked official first, then by how directly the repository is tied to the paper, then in the archive's order; never by how many ran. Licence labels name the repository's licence as recorded at harvest. “Pointer only” means Syntology does not serve that copy's text, for one of six reasons: no licence file was found; the licence was not identified; the licence is recorded as permissive but that copy's record is not marked cleared; the licence is outside the permissive list Syntology serves text under (MIT, Apache-2.0, BSD and similar); this paper's copy was not located, so its licence is not recorded; or this paper's copy is cleared, but the stored copy of the same code that Syntology serves is not. Some licences outside that list permit redistribution, such as WTFPL, and GPL-3.0 under its conditions; they are simply not on the list. Hover a licence label for the reason. File links open the file on GitHub at the default branch, which may have changed since the harvest.

A function or class name links to its page when the name occurs in the harvested code of at least 5 papers; the name is a string, not an identity. A census tag appears when Syntology's census ran implementations with that name on one shared input per bucket: it gives how many were compared, from how many papers, and how many distinct outputs they produced, and names this sample's bucket when this sample was one of them.

lucidrains/RETRO-pytorch community (archive-listed) 13 samples, 8 ran (2 instrument-failed) · 5 unverified
Attention in 344 papers
census: 13 implementations from 12 papers, 10 distinct outputs across 4 buckets, one shared input per bucket
retro_pytorch/retro_pytorch.py ran · metamorphic tier: deterministic Apache-2.0 (permissive) · 03b81e2f9a4d2d4c · report fixture: returned Attention · level recorded 2026-09-22
FeedForward in 198 papers retro_pytorch/retro_pytorch.py ran · metamorphic tier: deterministic Apache-2.0 (permissive) · 38be50d72d5874d6 · report fixture: returned FeedForward · level recorded 2026-09-22
PostNorm retro_pytorch/retro_pytorch.py ran · metamorphic tier: deterministic Apache-2.0 (permissive) · 562d2542a1eaacef · report fixture: returned PostNorm · level recorded 2026-09-22
PreNorm in 34 papers retro_pytorch/retro_pytorch.py ran · metamorphic tier: deterministic Apache-2.0 (permissive) · f44be7c955b39fa2 · report fixture: returned PreNorm · level recorded 2026-09-22
RMSNorm in 61 papers retro_pytorch/retro_pytorch.py ran · metamorphic tier: deterministic Apache-2.0 (permissive) · a4ae5f96225b96f3 · report fixture: returned RMSNorm · level recorded 2026-09-22
RotaryEmbedding in 27 papers retro_pytorch/retro_pytorch.py ran · metamorphic tier: well formed Apache-2.0 (permissive) · 9fbc41d552de00c6 · report fixture: signature max_seq_len:int/0/int, device:device/0/?, offset:int/0/int · returned Tensor · level recorded 2026-09-22
apply_rotary_pos_emb in 393 papers
census: 14 implementations from 15 papers, 8 distinct outputs across 7 buckets, one shared input per bucket this sample: bucket 1 of 7, arg 1: rank 3, kind float, dtype float32 · arg 2: rank 2, kind float, dtype float32
retro_pytorch/retro_pytorch.py ran · our draft was wrong Apache-2.0 (permissive) · a6120b0228dbf889 · report fixture: signature t:Tensor/3/float, freqs:Tensor/2/float · returned Tensor · level recorded 2026-09-22
rotate_half in 424 papers
census: 24 implementations from 96 papers, 4 distinct outputs across 2 buckets, one shared input per bucket this sample: bucket 1 of 2, arg 1: rank 2, kind float, dtype float32
retro_pytorch/retro_pytorch.py ran · our draft was wrong fingerprinted Apache-2.0 (permissive) · fdd93453f92f9167 · report fixture: signature x:Tensor/2/float · returned Tensor · level recorded 2026-09-22
ChunkedCrossAttention retro_pytorch/retro_pytorch.py unverified Apache-2.0 (permissive) · a18eb1addcd13eca · report
Decoder in 210 papers retro_pytorch/retro_pytorch.py unverified Apache-2.0 (permissive) · df5ec3eb8c24bd01 · report
Encoder in 278 papers retro_pytorch/retro_pytorch.py unverified Apache-2.0 (permissive) · dfd283b4f5317daf · report
RETRO retro_pytorch/retro_pytorch.py unverified Apache-2.0 (permissive) · c136963c983ee0fb · report
deepnorm_init retro_pytorch/retro_pytorch.py unverified Apache-2.0 (permissive) · 2f92b91f3350e0c9 · report
labmlai/annotated_deep_learning_paper_implementations community (archive-listed) 7 samples, 5 ran (0 instrument-failed) · 2 unverified
ChunkedCrossAttention labml_nn/transformers/retro/model.py ran · metamorphic tier: invariant MIT (permissive) · f393e5da1f357aea · report fixture: signature h:Tensor/3/float, e:Tensor/5/float · returned Tensor · level recorded 2026-09-22
CrossAttention in 85 papers labml_nn/transformers/retro/model.py ran · metamorphic tier: invariant MIT (permissive) · c3e56d6f2af858c1 · report fixture: signature e:Tensor/5/float, h:Tensor/4/float · returned Tensor · level recorded 2026-09-22
FeedForward in 198 papers labml_nn/transformers/retro/model.py ran · metamorphic tier: invariant MIT (permissive) · 1fae1fc6f059aacd · report fixture: signature h:Tensor/3/float · returned Tensor · level recorded 2026-09-22
RotaryPositionalEmbeddings labml_nn/transformers/retro/model.py ran · metamorphic tier: invariant fingerprinted MIT (permissive) · 6d63acdd4f116570 · report fixture: signature x:Tensor/4/float · returned Tensor · level recorded 2026-09-22
SelfAttention in 66 papers labml_nn/transformers/retro/model.py ran · metamorphic tier: invariant MIT (permissive) · 4c4fa4155f57fc48 · report fixture: signature h:Tensor/3/float · returned Tensor · level recorded 2026-09-22
NearestNeighborEncoder labml_nn/transformers/retro/model.py unverified MIT (permissive) · ed827370476041d7 · report
RetroModel labml_nn/transformers/retro/model.py unverified MIT (permissive) · 5156ed7e3c0f7fef · report
Repository not located 3 samples, 3 ran (0 instrument-failed) · 0 unverified
3 samples are identical to code Syntology harvested from another paper's repository; this paper's own copy was not located, so no repository is named and no licence is recorded.
cast_tuple in 58 papers repository not located ran · violated contract fingerprinted licence of this copy not recorded · 75e2de7f7799802d · report fixture: signature val:int/0/int, num:int/0/int · returned tuple · level recorded, date not recorded
default in 430 papers repository not located ran · violated contract fingerprinted licence of this copy not recorded · 60fff7c3c400d7ff · report fixture: signature val:int/0/int, d:int/0/int · returned int · level recorded, date not recorded
exists in 528 papers
census: 5 implementations from 103 papers, 1 distinct output on one shared input this sample: bucket 1 of 1, arg 1: rank 1, kind int, dtype int64
repository not located ran · violated contract licence of this copy not recorded · aa5486a3650902d8 · report fixture: signature val:Tensor/1/int · returned bool · level recorded, date not recorded

Method as code Syntology

What Syntology's graph records as this paper's claimed method, lifted into code: each assembly names its bricks (functions and classes), each with its own role and its own verification level. An assembly is not a check that the code implements the method correctly or reproduces the paper's results. Levels are each brick's own verification level in Syntology's graph (level 0: no recorded run), listed brick by brick and never combined into one level; “ran” means the brick executed on a synthesized input, not that its output is correct, and the instrument figure counts failures of Syntology's instrument, not of the code. Where a model read the code against the paper's claim, each assembly's label verdict is one model's judgement, kept in this page's JSON twin; it is not calibrated, so no card prints it. Siblings are assemblies lifted from this paper in the same batch for the same method: alternative readings kept side by side; none replaces or supersedes another.

2 assemblies on this paper: 2 harvested, 0 generated, listed in 1 groups (siblings lifted in one batch form one group).

2 sibling assemblies for Retrieval-Enhanced Transformer (R) (harvested), lifted in the same batch on 2026-09-09: alternative readings kept side by side; none replaces or supersedes another.

harvested Assembly for Retrieval-Enhanced Transformer (R) (graph method: Retrieval-Enhanced Transformer (R)). 16 bricks; individually on a fixture: 11 ran (of which 5 constructed an object rather than computing a result; 9 with no instrument failure: 0 honoured, 3 violated, 6 with no contract checked; 2 where Syntology's instrument failed) · 5 unverified. The assembly has not been run end-to-end.

one of 2 siblings

  1. RETRO entry · level 0 · unverified · c136963c983ee0fb
  2. Decoder model · level 0 · unverified · df5ec3eb8c24bd01
  3. Encoder model · level 0 · unverified · dfd283b4f5317daf
  4. RMSNorm helper · level 1 · ran · a4ae5f96225b96f3
  5. deepnorm_init helper · level 0 · unverified · 2f92b91f3350e0c9
  6. default helper · level 1 · ran · violated contract · 60fff7c3c400d7ff
  7. exists helper · level 1 · ran · violated contract · aa5486a3650902d8
  8. Attention model · level 1 · ran · 03b81e2f9a4d2d4c
  9. ChunkedCrossAttention model · level 0 · unverified · a18eb1addcd13eca
  10. FeedForward model · level 1 · ran · 38be50d72d5874d6
  11. PostNorm helper · level 1 · ran · 562d2542a1eaacef
  12. PreNorm helper · level 1 · ran · f44be7c955b39fa2
  13. RotaryEmbedding helper · level 1 · ran · 9fbc41d552de00c6
  14. apply_rotary_pos_emb helper · level 1 · ran · our draft was wrong · a6120b0228dbf889
  15. cast_tuple helper · level 1 · ran · violated contract · 75e2de7f7799802d
  16. rotate_half helper · level 1 · ran · our draft was wrong · fdd93453f92f9167

harvested Assembly for Retrieval-Enhanced Transformer (R) (graph method: Retrieval-Enhanced Transformer (R)). 7 bricks; individually on a fixture: 5 ran (of which 0 constructed an object rather than computing a result; 5 with no instrument failure: 0 honoured, 0 violated, 5 with no contract checked; 0 where Syntology's instrument failed) · 2 unverified. The assembly has not been run end-to-end.

one of 2 siblings

  1. RetroModel entry · level 0 · unverified · 5156ed7e3c0f7fef
  2. ChunkedCrossAttention model · level 1 · ran · f393e5da1f357aea
  3. FeedForward model · level 1 · ran · 1fae1fc6f059aacd
  4. NearestNeighborEncoder model · level 0 · unverified · ed827370476041d7
  5. SelfAttention model · level 1 · ran · 4c4fa4155f57fc48
  6. CrossAttention model · level 1 · ran · c3e56d6f2af858c1
  7. RotaryPositionalEmbeddings helper · level 1 · ran · 6d63acdd4f116570

Each brick ends with its code_sha256, Syntology's identity for that exact code; click one to copy the get_code call that returns it. For agents, on Syntology's MCP service: assembly(arxiv_id="2112.04426") returns these assemblies with each brick's role and wiring (how to connect).

Tasks

Language ModellingQuestion AnsweringRetrieval

Results from the paper archive 2025-07-28

TaskDatasetModelMetricValueRank at snapshotLeaderboardReport
Language Modelling WikiText-103 RETRO (7.5B) Number of params 7532M #1 of 89 Archive leaderboard report
Language Modelling WikiText-103 RETRO (7.5B) Test perplexity 2.4 #1 of 89 Archive leaderboard report
Question Answering Natural Questions RETRO + DPR (full) EM 45.5 #15 of 47 Archive leaderboard report

Ranks are positions in the archive's leaderboards as they stood at the 2025-07-28 snapshot. Results published since then are not among these rows, so a rank here is not a current standing.

Methods

Absolute Position EncodingsAdamAttentionAttention DropoutBERTBPECosine AnnealingDense ConnectionsDropoutGPT-3Label SmoothingLayer NormalizationLinear LayerLinear Warmup With Cosine AnnealingLinear Warmup With Linear DecayMulti-Head AttentionPosition-Wise Feed-Forward LayerResidual ConnectionSoftmaxTransformerWeight DecayWordPiece

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections