Retrieval-Augmented Generation for

Knowledge-Intensive NLP Tasks

Patrick Lewis†‡ , Ethan Perez? ,

arXiv:2005.11401v1 [cs.CL] 22 May 2020

Aleksandra Piktus† , Fabio Petroni† , Vladimir Karpukhin† , Naman Goyal† , Heinrich Küttler† ,

Mike Lewis† , Wen-tau Yih† , Tim Rocktäschel†‡ , Sebastian Riedel†‡ , Douwe Kiela†

†
Facebook AI Research; ‡ University College London; ? New York University;
plewis@fb.com

Abstract
Large pre-trained language models have been shown to store factual knowledge
in their parameters, and achieve state-of-the-art results when fine-tuned on down-
stream NLP tasks. However, their ability to access and precisely manipulate
knowledge is still limited, and hence on knowledge-intensive tasks, their perfor-
mance lags behind task-specific architectures. Additionally, providing provenance
for their decisions and updating their world knowledge remain open research prob-
lems. Pre-trained models with a differentiable access mechanism to explicit non-
parametric memory can overcome this issue, but have so far been only investigated
for extractive downstream tasks. We explore a general-purpose fine-tuning recipe
for retrieval-augmented generation (RAG) — models which combine pre-trained
parametric and non-parametric memory for language generation. We introduce
RAG models where the parametric memory is a pre-trained seq2seq model and
the non-parametric memory is a dense vector index of Wikipedia, accessed with
a pre-trained neural retriever. We compare two RAG formulations, one which
conditions on the same retrieved passages across the whole generated sequence, the
other can use different passages per token. We fine-tune and evaluate our models
on a wide range of knowledge-intensive NLP tasks and set the state-of-the-art
on three open domain QA tasks, outperforming parametric seq2seq models and
task-specific retrieve-and-extract architectures. For language generation tasks, we
find that RAG models generate more specific, diverse and factual language than a
state-of-the-art parametric-only seq2seq baseline.

1 Introduction
Pre-trained neural language models have been shown to learn a substantial amount of in-depth
knowledge from data [41]. They can do so without any access to an external memory, as a param-
eterized implicit knowledge base [45, 46]. While this development is exciting, such models do
have downsides: They cannot easily expand or revise their memory, can’t straightforwardly provide
insight into their predictions, and may produce “hallucinations” [34]. Hybrid models that combine
parametric memory with non-parametric (i.e., retrieval-based) memories [18, 22, 42] can address
some of these issues because knowledge can be directly revised and expanded, and its access can
be inspected and interpreted. REALM [18] and ORQA [27], two recently introduced models that
combine masked language models [8] with a differentiable retriever, have shown promising results,

Preprint. Under review.

 Define "middle ear"(x) The middle ear includes
End-to-End Backprop through q and pθ the tympanic cavity and
Question Answering: the three ossicles. (y)
Question Query Query Retriever pη Document Generator pθ Question Answering:
Answer Generation
Encoder (Non-Parametric)
Index
(Parametric)
Barack Obama was
d(z) supports (y)
born in Hawaii.(x) q(x) z4

Fact Verification: Fact Query z3 Margin- Fact Verification:

z2 Label Generation
alize
The Divine
Comedy (x) q MIPS z1 pθ This 14th century work
is divided into 3
Jeopardy Question sections: "Inferno",
Generation: "Purgatorio" &
Answer Query "Paradiso" (y)
Question Generation

Figure 1: An overview of retrieval-augmented generation (RAG). We combine a pre-trained re-

triever (Query Encoder + Document Index) with a pre-trained encoder-decoder (Generator) and
fine-tune end-to-end. For some query x, we use Maximum Inner Product Search (MIPS) to find the
top-K most relevant documents of all documents zi . To make the final prediction y, we treat z as a
latent variable and marginalize over the encoder-decoder predictions given different documents.

but have only explored open-domain extractive question answering. Here, we bring hybrid parametric
and non-parametric memory to the “workhorse of NLP,” i.e. sequence-to-sequence (seq2seq) models.
We endow pre-trained, parametric-memory generation models with a non-parametric memory through
a general-purpose fine-tuning approach which we refer to as retrieval-augmented generation (RAG).
We build RAG models where the parametric memory is a pre-trained generative seq2seq transformer,
and the non-parametric memory is a dense vector index of Wikipedia, accessed using a pre-trained
neural retriever. We combine these components in an end-to-end probabilistic model; the document
retriever (Dense Passage Retriever [22], henceforth DPR) provides latent documents conditioned on
the input, and the seq2seq model (BART [28]) then conditions on both these latent documents and
the input to generate the output. We marginalize the latent variables through a top-K approximation,
either on a per answer basis (assuming the same document is responsible for all tokens) or a per
answer token basis (assuming different documents can be responsible for different tokens). Just like
T5 [45] or BART, RAG can be fine-tuned on any seq2seq task, whereby both the sequence generator
and retriever are jointly learned.
There has been extensive previous work proposing architectures to enrich systems with non-parametric
memory which are trained from scratch for specific tasks—e.g. in memory networks [58, 49], stack-
augmented networks [21] and memory layers for transformers [26]. In contrast, we explore a setting
where both parametric and non-parametric memory components are pre-trained and pre-loaded with
extensive knowledge. Crucially, by using pre-trained knowledge-access mechanisms, the ability to
access knowledge is present without additional training.
Our results highlight the benefits of combining parametric and non-parametric memory with gen-
eration for knowledge-intensive tasks. Our RAG models achieve state-of-the-art results on open
Natural Questions [25], WebQuestions [3] and CuratedTrec [2] and strongly outperform recent
approaches that use specialised pre-training objectives on TriviaQA [20]. Despite these being ex-
tractive tasks, we find that unconstrained generation outperforms previous extractive approaches.
For knowledge-intensive generation, we experiment with MS-MARCO [1] and Jeopardy question
generation, and we find that our models generate responses that are more factual, specific, and diverse
than a BART baseline. For the FEVER [50] fact verification task, we achieve results within 4% of
sophisticated, state-of-the-art pipeline models which use strong supervision. Finally, we show that
the non-parametric memory can be replaced in order to control generation, demonstrating a simple
mechanism to update the knowledge that the model uses as facts about the world change.

2 Methods

We explore RAG models which use the input sequence x to retrieve text passages z and use these
passages as additional context when generating the target sequence y. As shown in Figure 1, our
models leverage two components: (i) a retriever pη (z|x) with parameters η that returns (top-K
truncated) distributions over text passages given a query x and (ii) a generator pθ (yi |x, z, y1:i−1 )

2
parametrized by θ that generates a current token based on a context of the previous i − 1 tokens
y1:i−1 , the original input x and a retrieved passage z.
To train the retriever and generator end-to-end, we treat the retrieved document as a latent variable.
We propose two models that marginalize over the latent documents in different ways to produce a
distribution over generated text. In one approach, RAG-Sequence, the model uses the same document
to predict each target token. In the other approach, RAG-Token, the model can predict each target
token based on a different document. In what follows, we formally introduce both models and then
describe the pη and pθ components, as well as the training and decoding procedure in more detail.

2.1 Models

RAG-Sequence Model The RAG-Sequence model uses the same retrieved document to generate
the complete sequence. Technically, it treats the retrieved passage as a single latent variable that is
marginalized to get the seq2seq probability p(y|x) via a top-K approximation,
X N
Y
pRAG-Sequence (y|x) = pη (z|x) pθ (yi |x, z, y1:i−1 ).
z∈top-k(p(·|x)) i

RAG-Token Model In the RAG-Token model we can draw a different latent passage for each
target token and marginalize accordingly. This allows the generator to choose content from several
documents when producing an answer. Formally, we define:
N
Y X
pRAG-Token (y|x) = pη (zi |x)pθ (yi |x, zi , y1:i−1 ).
i z∈top-k(p(·|x))

Finally, we note that RAG can be used for sequence classification tasks by considering the target class
as a target sequence of length one, in which case RAG-Sequence and RAG-Token are equivalent.

2.2 Retriever: DPR

The retrieval component pη (z|x) is based on DPR [22]. DPR follows a bi-encoder architecture:
pη (z|x) ∝ exp hd(z), q(x)i
where d(z) is a dense representation of the document produced by a BERTBASE transformer [8],
and q(x) a representation of the query by another BERTBASE transformer with a different set of
parameters.
To efficiently calculate top-k(pη (·|x)), the list of k elements z with highest prior probability pη (z|x),
DPR employs a Maximum Inner Product Search (MIPS) index provided by the FAISS library [19].
For non-parametric pre-trained memory, we use a pre-trained bi-encoder from [22] to both initialize
our retriever and to build the document index. This retriever was trained to retrieve documents which
contain answers to TriviaQA [20] questions and Natural Questions [25].

2.3 Generator: BART

The generator component pθ (yi |x, z, y1:i−1 ) could be modelled using any encoder-decoder. We use
BART-large [28], a pre-trained seq2seq transformer [52] with 400M parameters. To combine the
input x with the retrieved content z when generating from BART, we simply concatenate them.
BART was pre-trained using a denoising objective and a variety of different noising functions. It has
obtained state-of-the-art results on a diverse set of generation tasks and outperforms comparably-sized
T5 models [28]. We refer to the BART generator parameters θ as the parametric memory henceforth.

2.4 Training

We jointly train the retriever and generator components without any direct supervision on what
document should be retrieved. Given a fine-tuning training corpus of input/output pairs (xj , yj ), we

3
 P
minimize the negative marginal log-likelihood of each target, j − log p(yj |xj ) using stochastic
gradient descent with Adam [24]. Updating the document encoder during training is costly as it
requires the document index to be periodically updated as REALM does during pre-training [18].
We do not find this step necessary for strong performance, and we keep the document encoder (and
index) fixed, only fine-tuning the query encoder and the generator.

2.5 Decoding

At test/decoding time, RAG-Sequence and RAG-Token require different ways to approximate

arg maxy p(y|x).

RAG-Token The RAG-Token model can be seen as a standard, autoregressive, seq2seq generator
with transition probability:
X
p0θ (yi |x, y1:i−1 ) = pη (zi |x)pθ (yi |x, zi , y1:i−1 )
z∈top-k(p(·|x))

To decode, we can plug p0θ (yi |x, y1:i−1 ) into a standard beam decoder.

RAG-Sequence The likelihood p(y|x) does not break into a conventional per-token likelihood for
the RAG-Sequence, and hence we cannot solve it with a single beam search pass. Instead, we run
beam search for each candidate document z, scoring each hypothesis using pθ (yi |x, z, y1:i−1 ). This
yields a set of hypotheses Y of which some might not have appeared in the beams of all documents.
To estimate the probability of an hypothesis y across all beams, we run an additional forward pass
for each document z for which y does not appear in the beam, multiply the generator score with
pη (z|x) and then sum up the probabilities across beams for the marginals. We refer to this decoding
procedure as “Thorough Decoding.”
For longer output sequences, |Y | can become large, requiring many forward passes. For more efficient
decoding, we can make a further approximation that pθ (y|x, zi ) ≈ 0 where y was not generated
during beam search from x, zi . This avoids the need to run additional forward passes once the
candidate set Y has been generated. We refer to this decoding procedure as “Fast Decoding”.

3 Experiments
We experiment with RAG in a wide range of knowledge-intensive tasks. For all experiments, we
use a single Wikipedia dump for our non-parametric knowledge source. Following Lee et al. [27]
and Karpukhin et al. [22], we use the December 2018 dump. Each Wikipedia article is split into
disjoint 100-word chunks, to make a total of 21,015,324 documents.1 We use the DPR document
encoder to compute document embeddings for each document, and we build a single MIPS index
using FAISS [19] using Hierarchical Navigable Small World approximation for efficient retrieval [33],
which is then used for all experiments. During training, we retrieve the top k documents for each
query, where we consider k ∈ {5, 10}. We determine k for test time using validation data. In the
remainder of this section, we will discuss the experimental details for each of these task settings.

3.1 Open-domain Question Answering

Open-domain QA is an important real-world NLP application and is often used as test-bed for
knowledge-intensive tasks [18]. We tackle open-domain QA by treating questions and answers as
simple input-output text pairs (x, y), and we train RAG by directly minimizing the negative log-
likelihood of answers. We compare our results to the popular extractive QA paradigm [5, 7, 27, 22],
where answers are extracted as spans from retrieved documents, relying primarily on non-parametric
knowledge. In addition, we also compare to "Closed-Book QA" approaches [46], which, like RAG,
generate answers, but do not exploit latent retrieval, instead relying purely on parametric knowledge.
We consider four popular open-domain QA datasets: Natural Questions (NQ) [25], TriviaQA
(TQA) [20]. WebQuestions (WQ) [3] and CuratedTrec (CT) [2]. The answers for CuratedTrec
are given in the form of regular expressions, which has been cited as a reason why it is unsuitable
1
The reader is referred to Karpukhin et al. [22] for further details on how Wikipedia is pre-processed.

4
for answer-generation models [18]. To overcome this, we use a pre-processing step where we first
retrieve the top 1000 documents for each query, and use the answer that most frequently matches
the regex pattern as the supervision target. If no matches are found, we resort to a simple heuristic:
generate all possible permutations for each regex, replacing non-deterministic symbols in the regex
nested tree structure with a whitespace. As CuratedTrec and WebQuestions are small datasets, we
follow DPR [22] by initializing CuratedTrec and WebQuestions models with our Natural Questions
RAG model.
We use the same training/dev/testing splitting method as in previous work [27, 22] and report the
standard Exact Match (EM) metric. For TriviaQA, in order to compare to T5 [46], we do an additional
test evaluation on the TriviaQA Wiki test set.

3.2 Abstractive Question Answering

Because RAG leverages an encoder-decoder model, it can go beyond extractive question-answering

and answer questions with free-form, abstractive text generation. To test RAG’s ability to generate
natural language responses in a knowledge-intensive setting, we use the MS-MARCO Natural
Language Generation task v2.1 [38]. This task consists of natural language questions submitted to
a search engine, ten snippets retrieved from a search engine for each question, and a full sentence
natural language answer annotated from these retrieved passages.
As we are interested in models that can perform their own latent retrieval, we do not use the supplied
passages, only the questions and answers, thus treating MS-MARCO as an open-domain abstractive
question answering task. MS-MARCO does contain some questions that cannot be answered in a
way that matches the reference answer without access to the context passages, such as “What is the
weather in volcano, CA?” so we note that performance on Open-MSMARCO will be lower than
models that do use these gold context passages.
We further note that there are questions in MS-MARCO that cannot be answered using a Wikipedia
knowledge source alone. In these cases, RAG can rely on the parametric implicit knowledge in its
BART parameters in order to generate commonsense responses.

3.3 Jeopardy Question Generation

In order to further evaluate RAG’s generation abilities in a non-question answering setting, we

propose to study Open-domain question generation. Rather than repurpose questions from standard
open-domain QA tasks, which typically consist of short and simple questions, we instead propose to
study the more demanding task of generating of Jeopardy questions. Jeopardy is an unusual format
that consists of trying to guess an entity from a fact about that entity. For example, “The World
Cup” is the answer to the jeopardy question “In 1986 Mexico scored as the first country to host this
international sports competition twice.” As Jeopardy “questions" are precise, factual statements,
generating Jeopardy-style questions conditioned on the answer entity they refer to constitutes a
challenging knowledge-intensive generation task.
We use the raw Jeopardy data and splits from SearchQA [10], consisting of 97,391 training, 13,713
development, and 26,848 test datapoints. As this is a new task, we also train a BART system to
compare RAG to. Following [60], we evaluate generations using the SQuAD-tuned Q-BLEU-1
metric [37]. Q-BLEU-1 is a variant of BLEU-1 which puts a higher weight on the matching entities,
and has higher correlation with human judgment for question generation compared to standard word
overlap metrics.
As automatic metrics can be unreliable, especially on such open-ended tasks, we also perform a
human evaluation of generations. We run two evaluations, one to assess the factuality of generations,
and one to assess specificity. We follow the recent best-practice of performing a pairwise comparative
evaluation between two systems [30]. Assessors are shown an answer entity and two generated
questions about that entity, one from BART and one from RAG. They are then asked to pick one of
four possible options—Sentence A is better, Sentence B is better, both are correct or neither is good.

3.4 Fact Verification

FEVER [50] is a fact verification dataset that involves classifying whether a natural language claim
is supported or refuted by Wikipedia, or whether there is not enough information to decide. The

5
 Model NQ TQA WQ CT
T5-11B [46] 34.5 - /50.1 37.4 -
Closed-Book
T5-11B + SSM [46] 36.6 - /60.5 44.7 -
REALM [18] 40.4 - / - 40.7 46.8
Open-Book
DPR [22] 41.5 57.9/ - 41.1 50.6
RAG-Token 44.1 55.2/66.1 45.5 50.0
RAG-Sequence 44.5 56.1/68.0 45.2 52.2

Table 1: Open-Domain QA Test Scores. For TQA, the left column uses the test split commonly used
in Open-Domain QA. The right column uses the hidden TQA Wiki test split. See Appendix B for
further information.

Model Jeopardy QGen MS-MARCO FEVER-3 FEVER-2

B-1 QB-1 R-L B-1 Label Accuracy
SotA - - 49.8* 49.9* 76.8 92.2*
BART 15.1 19.7 38.2 41.6 64.0 81.1
RAG-Token 17.3 22.2 40.1 41.5
72.5 89.5
RAG-Sequence 14.7 21.4 40.8 44.2

Table 2: Generation and classification task Test Scores. SotA for MS-MARCO is [4], FEVER-3 is
[61] and FEVER-2 is [51] * Uses gold context/evidence, best-performing model without gold access
underlined. As FEVER is a classification dataset, RAG-Token and RAG-Sequence are equivalent.

task requires retrieving evidence from Wikipedia relating to the claim and then reasoning about the
retrieved evidence to classify whether the claim is true, false, or unverifiable from Wikipedia alone.
FEVER is a retrieval problem coupled with an entailment reasoning task. It also provides a good test
bed for exploring the RAG models’ ability to handle classification rather than generation.
We map FEVER class labels (supports, refutes, or not enough info) to single output tokens and
directly train with claim-class pairs. Crucially, unlike most other approaches to FEVER, we do not
use supervision on retrieved evidence. We explore two different FEVER variants: the standard 3-way
classification task (supports/refutes/not enough info) and the 2-way FEVER (supports/refutes) task
studied in Thorne and Vlachos [51]. In both cases we report label accuracy.

3.5 Implementation Details

For Open-domain QA we report test numbers using 15 retrieved documents for RAG-Token models.
For RAG-Sequence models, we report test results using 50 retrieved documents, and we use the
Thorough Decoding approach since answers are generally short. We use greedy decoding for QA as
we did not find beam search improved results. For Open-MSMarco and Jeopardy question generation,
we report test numbers using ten retrieved documents for both RAG-Token and RAG-Sequence, and
we also train a BART-large model as a baseline. We use a beam size of four, and use the Fast Decoding
approach for RAG-Sequence models, as Thorough Decoding did not improve performance.

4 Results
4.1 Open-domain Question Answering

Table 1 shows results for RAG along with recent state-of-the-art models. On all four open-domain
QA tasks, RAG sets a new state-of-the-art (in the case of TQA only on the T5-comparable split).
RAG combines the generation flexibility of the “closed-book" (parametric only) approaches and the
performance of "open-book" retrieval-based approaches. Unlike REALM and T5+SSM, RAG enjoys
strong results without expensive specialized "salient span masking" pre-training [18], relying on
off-the-shelf components. It is worth noting that RAG’s retriever is initialized using DPR’s retriever,

6
 BART better RAG-Token better Both good Both poor No Majority
Factuality 7.1% 42.7% 11.7% 17.7% 20.8%
Specificity 16.8% 37.4% 18.8% 6.9% 20.1%

Table 3: Human assessments for the Jeopardy Question Generation Task.

which does use retrieval supervision on Natural Questions and TriviaQA. RAG compares favourably
to DPR QA system on open-domain QA, which uses a BERT-based cross-encoder system to re-rank
documents, along with an extractive reader. RAG demonstrates that neither a re-ranker nor extractive
reader is necessary for state-of-the-art machine reading performance. Generating answers even when
it is possible to extract them has a number of advantages. Documents which contain clues as to
the correct answer but do not contain the correct answer verbatim themselves can still contribute
probability mass towards a correct answer being generated, which is not possible with standard
extractive approaches, leading to more effective marginalization across documents. Furthermore,
RAG is able to generate correct answers even when the correct answer is not present in any of the
retrieved documents, achieving an accuracy of 11.8% in such cases for Natural Questions, whereas
an extractive model would score 0%.

4.2 Abstractive Question Answering

As shown in Table 2, RAG-Sequence outperforms BART on Open MS-MARCO generation by

2.6 Bleu points and 2.6 Rouge-L points. It approaches the performance of state-of-the-art models,
which is impressive considering that (i) these models have access to passages that contain the
specific information required to generate the reference answer, (ii) many questions are unanswerable
without access to gold passages, and (iii) other questions are unanswerable from Wikipedia alone.
Table 4 shows some generated answers from our models. Qualitatively, we find that RAG models
hallucinate less and generate factually correct text more often than BART. Later we also show that
RAG generations are more diverse than BART generations (see Section 4.6).

4.3 Jeopardy Question Generation

Table 2 shows automatic metric results on the Jeopardy question generation task. We find that RAG-
Token performs better than the RAG-Sequence model in this setting, with both models outperforming
BART using the Q-BLEU-1 metric.
Table 3 shows the results from the human evaluation. The human evaluation was carried out with
452 pairs of generations from BART and RAG-Token. The annotators indicated that BART was
more factual than RAG in only 7.1% of cases, while RAG was more factual in 42.7% of cases and
both RAG and BART were factual in a further 17% of cases, clearly demonstrating the comparative
effectiveness of RAG on the task over a state-of-the-art conditional generation model. The annotators
also strongly prefer RAG generations in terms of specificity.
Typical example of generations from each model are shown in Table 4. BART generates a more
generic response (which is incorrect), whereas the RAG models generate specific and correct facts
about Washington state.
We hypothesise that RAG-Token performs best for this task as Jeopardy questions often contain two
separate pieces of information about the entity, and RAG-Token is able to synthesize a response by
combining disparate information from different retrieved documents in one generation. Figure 2
shows an example where content from two documents has been combined to produce the generated
question. Document 2 contains information about Hemingway’s “The Sun also rises,” and the
contribution for “Sun” is very high for document 2. Similarly, “A Farewell to Arms” is mentioned
in Document 1, which dominates the posterior when this title is generated. Intriguingly, after the
first token of these book titles are generated, the distribution over documents flattens again. This
observation suggests that the generator completes the book titles without depending on specific
documents. In other words, the model’s parametric knowledge is sufficient to complete the titles.
We show evidence for the above interpretation by feeding the BART-only baseline with the partial
decoding "The Sun. BART completes the generation "The Sun Also Rises" is a novel by this

7
Document 1: his works are considered classics of American Doc 1
literature ... His wartime experiences formed the basis for his novel Doc 2
”A Farewell to Arms” (1929) ...
Doc 3
Document 2: ... artists of the 1920s ”Lost Generation” expatriate
Doc 4
community. His debut novel, ”The Sun Also Rises”, was published
in 1926. Doc 5

S
”
e
n

by

”
A
so

is

au s

of
or

re

”
ell
to

s
ve
ise

m
Th
BO

Su

th

Fa
th
Al

Ar
no
Figure 2: RAG-Token document posterior p(zi |x, yi , y−i ) for each generated token for input “Hem-
ingway" for Jeopardy generation with 5 retrieved documents. The posterior for document 1 is high
when generating “A Farewell to Arms" and for document 2 when generating “The Sun Also Rises"

Task Input Model Generation

?
BART The middle ear is the part of the ear between the middle ear and the nose.
define middle
RAG-T The middle ear is the portion of the ear internal to the eardrum.
ear
MS- RAG-S The middle ear includes the tympanic cavity and the three ossicles.
MARCO
what currency BART The currency needed in Scotland is Pound sterling.
needed in RAG-T Pound is the currency needed in Scotland.
scotland RAG-S The currency needed in Scotland is the pound sterling.
?
BART This state has the largest number of counties in the U.S.
Jeopardy Washington RAG-T It’s the only U.S. state named for a U.S. president
Question RAG-S It’s the state where you’ll find Mount Rainier National Park
Gener *
-ation BART This epic poem by Dante is divided into 3 parts: the Inferno, the Purgatorio & the Purgatorio
The Divine
RAG-T Dante’s "Inferno" is the first part of this epic poem
Comedy
RAG-S This 14th century work is divided into 3 sections: "Inferno", "Purgatorio" & "Paradiso"

Table 4: Example Generations for MS-MARCO and Jeopardy Question generation. RAG models
generate mpre specific and factually accurate responses, whereas BART generate more factually
incorrect (marked by ‘?’), or partially correct (marked by *) and more generic responses.

author of "The Sun Also Rises" indicating the title "The Sun Also Rises" is stored in BART’s
parameters. Similarly, feeding the partial decoding "The Sun Also Rises" is a novel by this
author of "A will result in BART completing the generation with "The Sun Also Rises" is a
novel by this author of "A Farewell to Arms. This example shows how the parametric and non-
parametric memories work together—the non-parametric component helps to guide the generation in
a particular direction, drawing out specific knowledge stored in the parametric memory.

4.4 Fact Verification

Table 2 shows our results on the FEVER 3-way and 2-way classification task. For 3-way classification,
RAG achieves accuracies that are within 4.3% of state-of-the-art models, which are complex pipeline
systems with domain-specific architectures and substantial engineering, trained using intermediate
supervision, which RAG does not require.
For 2-way classification, we compare against the model from [51], which trains RoBERTa [31] to
classify the claim as true or false given the gold evidence sentence. RAG achieves an accuracy within
2.7% of this model, despite being supplied with only the claim and retrieving its own evidence.
We also analyze whether the documents retrieved by RAG correspond to the documents annotated as
gold evidence in FEVER. We analyze the overlap in Wikipedia articles between the top k documents
retrieved by RAG and the gold, annotated evidence documents. We find that the top article retrieved
by RAG is a gold document for the claim in 71% of cases, and a gold article is present in the top 10
retrieved articles in 90% of cases.

4.5 Ablations

To gain a better understanding of what factors affect RAG’s performance, we perform a number of
ablation experiments for our tasks on their respective development sets.

8
 Model NQ TQA WQ CT Jeopardy-QGen MSMarco FVR-3 FVR-2
Exact Match B-1 QB-1 R-L B-1 Label Accuracy
RAG-Token-BM25 29.7 41.5 32.1 33.1 17.5 22.3 55.5 48.4
75.1 91.6
RAG-Seq-BM25 31.8 44.1 36.6 33.8 11.1 19.5 56.5 46.9
RAG-Token-Frozen 37.8 50.1 37.1 51.1 16.7 21.7 55.9 49.4
72.9 89.4
RAG-Seq-Frozen 41.2 52.1 41.8 52.6 11.8 19.6 56.7 47.3
RAG-Token 43.5 54.8 46.5 51.9 17.9 22.6 56.2 49.4
74.5 90.6
RAG-Seq 44.0 55.8 44.9 53.4 15.3 21.5 57.2 47.5

Table 5: Ablations on the development set. As FEVER is a classification dataset, RAG-Token and
RAG-Sequence are equivalent.

44 80
NQ Answer Recall @ K

Bleu-1 / Rouge-L score

56
43
NQ Exact Match

70
54 RAG-Tok R-L
42 RAG-Tok B-1
60
52 RAG-Seq R-L
41 RAG-Tok
RAG-Seq B-1
50 RAG-Seq 50
40
RAG-Tok Fixed DPR
RAG-Seq 40 BM25 48
39

10 20 30 40 50 10 20 30 40 50 10 20 30 40 50
K Retrieved Docs K Retrieved Docs K Retrieved Docs

Figure 3: Left: NQ performance as more documents are retrieved. Center: Fraction of answers in
NQ where the answer occurs somewhere in the top K documents. Right: MS-MARCO Bleu-1 and
Rouge-L as more documents are retrieved.

Using more documents Models are trained with either 5 or 10 retrieved latent documents, and we
do not observe significant differences in performance between them. We also have the flexibility to
adjust the number of retrieved documents at test time, which does affect performance. Figure 3 (left)
shows that retrieving more documents at test time monotonically improves Open-domain QA results
for RAG-Sequence, but performance peaks for RAG-Token at 10 retrieved documents. Figure 3
(right) shows that retrieving more documents leads to higher Rouge-L for RAG-Token at the expense
of Bleu-1, but the effect is less pronounced for RAG-Sequence.

Retrieval A key feature of RAG is the ability to learn to retrieve relevant information for the task
at hand. To assess the effectiveness of the retrieval mechanism, we run ablations on RAG where we
prevent gradients from propagating into the retriever. Table 5 shows the results across all tasks. In
each case, learned retrieval improves results, with the largest improvements in question answering.
Figure 3 (center) shows that the learned retriever shows a higher recall for gold documents compared
to the fixed retriever. The improvements on TriviaQA and Natural Questions are notable, as we
initialize the retriever from DPR, which is trained with strong, document-level supervision to perform
well on these tasks. We also compare RAG’s dense embedding-based retrieval mechanism to a word
overlap-based BM25 retriever [47]. Here, we replace RAG’s differentiable retriever with a fixed
BM25 system. We use the BM25 retrieval scores as logits when calculating pη (zi |x). Table 5 and
Figure 3 show the results. For FEVER, we find that BM25 performs best, perhaps since FEVER
claims are heavily entity-centric and thus well-suited for word overlap-based retrieval. On all other
tasks, we find the differentiable retrieval to be helpful, especially question answering, where it is
crucial.

4.6 Generation Diversity

Section 4.3 established that RAG models generate are more factual and specific than BART for
Jeopardy question generation. Similar to Li et al. [29], Vijayakumar et al. [53] and Massarelli et al.
[35], we also investigate the diversity of generations by calculating the ratio of distinct ngrams to
total ngrams generated by different models. Table 6 shows that RAG-Sequence generations are more

9
 Dataset Gold BART RAG-Token RAG-Sequence
MSMARCO 89.6% 70.7% 77.8% 83.5%
Jeopardy Generation 90.0% 32.4% 46.8 % 53.8%

Table 6: Ratio of distinct tri-grams to total tri-grams in the development set generations for MS-
MARCO and Jeopardy Question Generation.

diverse than RAG-Token generations, and both generate significantly more diverse outputs than
BART without requiring any diversity-promoting decoding strategy.

4.7 Hot-swapping indices

An advantage of non-parametric knowledge models such as RAG is that the knowledge base can be
easily updated at test time. Parametric-only models such as T5 or BART require additional training to
update their behavior as facts about the world change. As a demonstration, we build an index using
the DrQA Wikipedia dump [5], (dated December 21st, 2016) and compare generations from RAG
using this index to the newer index used in our main results (December 20th, 2018). We prepared a list
of 82 heads of states who had changed between these dates and used a template “Who is {position}?”
(e.g., “Who is the prime minister of the UK?”) to query our Natural Questions -finetuned RAG model
with each index. RAG achieved an accuracy of 70% using the 2016 index for 2016 world leaders
and an accuracy of 68% using the 2018 index for the 2018 world leaders. Only 21% of the model’s
predictions were the same using the two indices, and accuracy using mismatched indices is very low
(12% using the 2018 index for 2016 leaders and 4% using the 2016 index for 2018 leaders). Our
result shows that we can effectively update RAG’s behavior with new world knowledge by simply
replacing its non-parametric memory.

5 Related Work
Single-Task Retrieval Prior work has shown that retrieval improves performance across a variety of
NLP tasks when considered in isolation. Such tasks include open-domain question answering [5, 25],
fact checking [50], fact completion [42], long-form question answering [12], Wikipedia article
generation [32], dialogue [36, 59, 9, 13], translation [16], and language modeling [17, 23]. Our
work unifies previous successes in incorporating retrieval into individual tasks, showing that a single
retrieval-based architecture is capable of achieving strong performance across several tasks.

General-Purpose Architectures for NLP Prior work on general-purpose architectures for NLP
tasks has shown great success without the use of retrieval. A single, pre-trained language model
has been shown to achieve strong performance on various classification tasks in the GLUE bench-
marks [54, 55] after fine-tuning [43, 8]. GPT-2 [44] later showed that a single, left-to-right, pre-trained
language model could achieve strong performance across both discriminative and generative tasks.
For further improvement, BART [28] and T5 [45, 46] propose a single, pre-trained encoder-decoder
model that leverages bi-directional attention to achieve stronger performance on discriminative
and generative tasks. Our work aims to expand the space of possible tasks with a single, unified
architecture, by learning a retrieval module to augment pre-trained, generative language models.

Learned Retrieval There is significant work on learning to retrieve documents in information

retrieval, more recently with pre-trained, neural language models [39, 22] similar to ours. Some
work optimizes the retrieval module to aid in a specific, downstream task such as question answering,
using search [40], reinforcement learning [6, 57, 56], or a latent variable approach [27, 18] as in our
work. These successes leverage different retrieval-based architectures and optimization techniques to
achieve strong performance on a single task, while we show that a single retrieval-based architecture
can be fine-tuned for strong performance on a variety of tasks.

Memory-based Architectures Our document index can be seen as a large external memory for
neural networks to attend to, analagous to memory networks [58, 48]. Concurrent work [14] learns
to retrieve a trained embedding for each entity in the input, rather than to retrieve raw text as in our

10
work. Other work improves the ability of dialog models to generate factual text by attending over
fact embeddings [9, 13] or, closer to our work, over retrieved text directly [15]. A key feature of our
memory is that it is comprised of raw text rather distributed representations, which makes the memory
both (i) human-readable, lending a form of interpretability to our model, and (ii) human-writable,
enabling us to dynamically update the model’s memory by editing the document index.

6 Discussion
In this work, we presented hybrid generation models with access to parametric and non-parametric
retrieval-based external memory, in the form of Wikipedia. We showed that our RAG models obtain
state-of-the-art performance on open domain question answering. We found that people prefer RAG’s
generation over purely parametric BART and find RAG more factual, and we conducted a detailed
investigation of the learned retrieval component, validating its effectiveness. We also showed that
the model’s grounding in external data leads it to generate more diverse, and illustrated by how the
retrieval index can be hot-swapped on the fly without having to retrain the model. In future work,
it would be interesting to investigate if the two components can be jointly pre-trained from scratch,
either on a denoising objective similar to BART, or through some other objective. Our work opens
new research directions on how parametric and non-parametric memories interact and how to most
effectively combine the different components, showing promise in being applied to a wide variety of
NLP tasks.

Acknowledgments
EP thanks supports from the NSF Graduate Research Fellowship.

References
[1] Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan
Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, Mir Rosenberg, Xia Song, Alina
Stoica, Saurabh Tiwary, and Tong Wang. MS MARCO: A Human Generated MAchine
Reading COmprehension Dataset. arXiv:1611.09268 [cs], November 2016. URL http:
//arxiv.org/abs/1611.09268. arXiv: 1611.09268.

[2] Petr Baudiš and Jan Šedivỳ. Modeling of the question answering task in the yodaqa system. In
International Conference of the Cross-Language Evaluation Forum for European Languages,
pages 222–228. Springer, 2015. URL https://link.springer.com/chapter/10.1007%
2F978-3-319-24027-5_20.
[3] Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang. Semantic Parsing on Freebase
from Question-Answer Pairs. In Proceedings of the 2013 Conference on Empirical Methods
in Natural Language Processing, pages 1533–1544, Seattle, Washington, USA, October 2013.
Association for Computational Linguistics. URL http://www.aclweb.org/anthology/
D13-1160.
[4] Bin Bi, Chenliang Li, Chen Wu, Ming Yan, and Wei Wang. Palm: Pre-training an autoencod-
ing&autoregressive language model for context-conditioned generation. ArXiv, abs/2004.07159,
2020. URL https://arxiv.org/abs/2004.07159.
[5] Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. Reading Wikipedia to Answer
Open-Domain Questions. In Proceedings of the 55th Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers), pages 1870–1879, Vancouver, Canada,
July 2017. Association for Computational Linguistics. doi: 10.18653/v1/P17-1171. URL
https://www.aclweb.org/anthology/P17-1171.
[6] Eunsol Choi, Daniel Hewlett, Jakob Uszkoreit, Illia Polosukhin, Alexandre Lacoste, and
Jonathan Berant. Coarse-to-fine question answering for long documents. In Proceedings of the
55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),
pages 209–220, Vancouver, Canada, July 2017. Association for Computational Linguistics. doi:
10.18653/v1/P17-1020. URL https://www.aclweb.org/anthology/P17-1020.

11
 [7] Christopher Clark and Matt Gardner. Simple and Effective Multi-Paragraph Reading Compre-
hension. arXiv:1710.10723 [cs], October 2017. URL http://arxiv.org/abs/1710.10723.
arXiv: 1710.10723.
[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of
Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Con-
ference of the North American Chapter of the Association for Computational Linguistics: Human
Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis,
Minnesota, June 2019. Association for Computational Linguistics. doi: 10.18653/v1/N19-1423.
URL https://www.aclweb.org/anthology/N19-1423.
[9] Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. Wiz-
ard of wikipedia: Knowledge-powered conversational agents. In International Conference on
Learning Representations, 2019. URL https://openreview.net/forum?id=r1l73iRqKm.
[10] Matthew Dunn, Levent Sagun, Mike Higgins, V. Ugur Guney, Volkan Cirik, and Kyunghyun
Cho. SearchQA: A New Q&A Dataset Augmented with Context from a Search Engine.
arXiv:1704.05179 [cs], April 2017. URL http://arxiv.org/abs/1704.05179. arXiv:
1704.05179.
[11] Angela Fan, Mike Lewis, and Yann Dauphin. Hierarchical neural story generation. In Proceed-
ings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1:
Long Papers), pages 889–898, Melbourne, Australia, July 2018. Association for Computational
Linguistics. doi: 10.18653/v1/P18-1082. URL https://www.aclweb.org/anthology/
P18-1082.
[12] Angela Fan, Yacine Jernite, Ethan Perez, David Grangier, Jason Weston, and Michael Auli. ELI5:
Long form question answering. In Proceedings of the 57th Annual Meeting of the Association
for Computational Linguistics, pages 3558–3567, Florence, Italy, July 2019. Association for
Computational Linguistics. doi: 10.18653/v1/P19-1346. URL https://www.aclweb.org/
anthology/P19-1346.
[13] Angela Fan, Claire Gardent, Chloe Braud, and Antoine Bordes. Augmenting transformers
with KNN-based composite memory, 2020. URL https://openreview.net/forum?id=
H1gx1CNKPH.
[14] Thibault Févry, Livio Baldini Soares, Nicholas FitzGerald, Eunsol Choi, and Tom Kwiatkowski.
Entities as experts: Sparse memory access with entity supervision. ArXiv, abs/2004.07202,
2020. URL https://arxiv.org/abs/2004.07202.
[15] Marjan Ghazvininejad, Chris Brockett, Ming-Wei Chang, Bill Dolan, Jianfeng Gao, Wen
tau Yih, and Michel Galley. A knowledge-grounded neural conversation model. In AAAI
Conference on Artificial Intelligence, 2018. URL https://www.aaai.org/ocs/index.php/
AAAI/AAAI18/paper/view/16710.
[16] Jiatao Gu, Yong Wang, Kyunghyun Cho, and Victor O.K. Li. Search engine guided neural
machine translation. In AAAI Conference on Artificial Intelligence, 2018. URL https:
//www.aaai.org/ocs/index.php/AAAI/AAAI18/paper/view/17282.
[17] Kelvin Guu, Tatsunori B. Hashimoto, Yonatan Oren, and Percy Liang. Generating sentences by
editing prototypes. Transactions of the Association for Computational Linguistics, 6:437–450,
2018. doi: 10.1162/tacl_a_00030. URL https://www.aclweb.org/anthology/Q18-1031.
[18] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. REALM:
Retrieval-augmented language model pre-training. ArXiv, abs/2002.08909, 2020. URL https:
//arxiv.org/abs/2002.08909.
[19] Jeff Johnson, Matthijs Douze, and Hervé Jégou. Billion-scale similarity search with gpus. arXiv
preprint arXiv:1702.08734, 2017. URL https://arxiv.org/abs/1702.08734.
[20] Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. TriviaQA: A Large Scale
Distantly Supervised Challenge Dataset for Reading Comprehension. In Proceedings of the
55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),

12
 pages 1601–1611, Vancouver, Canada, July 2017. Association for Computational Linguistics.
doi: 10.18653/v1/P17-1147. URL https://www.aclweb.org/anthology/P17-1147.
[21] Armand Joulin and Tomas Mikolov. Inferring algorithmic patterns with stack-
augmented recurrent nets. In Proceedings of the 28th International Conference on
Neural Information Processing Systems - Volume 1, NIPS’15, page 190–198, Cam-
bridge, MA, USA, 2015. MIT Press. URL https://papers.nips.cc/paper/
5857-inferring-algorithmic-patterns-with-stack-augmented-recurrent-nets.
[22] Vladimir Karpukhin, Barlas Oguz, Sewon Min, Ledell Wu, Sergey Edunov, Danqi Chen, and
Wen-tau Yih. Dense passage retrieval for open-domain question answering. arXiv preprint
arXiv:2004.04906, 2020. URL https://arxiv.org/abs/2004.04906.
[23] Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. Generaliza-
tion through memorization: Nearest neighbor language models. In International Conference on
Learning Representations, 2020. URL https://openreview.net/forum?id=HklBjCEKvH.
[24] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua
Bengio and Yann LeCun, editors, 3rd International Conference on Learning Representations,
ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. URL
http://arxiv.org/abs/1412.6980.
[25] Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh,
Chris Alberti, Danielle Epstein, Illia Polosukhin, Matthew Kelcey, Jacob Devlin, Ken-
ton Lee, Kristina N. Toutanova, Llion Jones, Ming-Wei Chang, Andrew Dai, Jakob
Uszkoreit, Quoc Le, and Slav Petrov. Natural Questions: a Benchmark for Ques-
tion Answering Research. Transactions of the Association of Computational Lin-
guistics, 2019. URL https://tomkwiat.users.x20web.corp.google.com/papers/
natural-questions/main-1455-kwiatkowski.pdf.
[26] Guillaume Lample, Alexandre Sablayrolles, Marc’ Aurelio Ranzato, Ludovic Denoyer, and
Herve Jegou. Large memory layers with product keys. In H. Wallach, H. Larochelle,
A. Beygelzimer, F. d’ Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural In-
formation Processing Systems 32, pages 8548–8559. Curran Associates, Inc., 2019. URL http:
//papers.nips.cc/paper/9061-large-memory-layers-with-product-keys.pdf.
[27] Kenton Lee, Ming-Wei Chang, and Kristina Toutanova. Latent retrieval for weakly supervised
open domain question answering. In Proceedings of the 57th Annual Meeting of the Association
for Computational Linguistics, pages 6086–6096, Florence, Italy, July 2019. Association for
Computational Linguistics. doi: 10.18653/v1/P19-1612. URL https://www.aclweb.org/
anthology/P19-1612.
[28] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed,
Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. BART: Denoising sequence-to-sequence
pre-training for natural language generation, translation, and comprehension. arXiv preprint
arXiv:1910.13461, 2019. URL https://arxiv.org/abs/1910.13461.
[29] Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. A diversity-promoting
objective function for neural conversation models. In Proceedings of the 2016 Conference of the
North American Chapter of the Association for Computational Linguistics: Human Language
Technologies, pages 110–119, San Diego, California, June 2016. Association for Computational
Linguistics. doi: 10.18653/v1/N16-1014. URL https://www.aclweb.org/anthology/
N16-1014.
[30] Margaret Li, Jason Weston, and Stephen Roller. Acute-eval: Improved dialogue evaluation
with optimized questions and multi-turn comparisons. ArXiv, abs/1909.03087, 2019. URL
https://arxiv.org/abs/1909.03087.
[31] Hairong Liu, Mingbo Ma, Liang Huang, Hao Xiong, and Zhongjun He. Robust neural machine
translation with joint textual and phonetic embedding. In Proceedings of the 57th Annual
Meeting of the Association for Computational Linguistics, pages 3044–3049, Florence, Italy,
July 2019. Association for Computational Linguistics. doi: 10.18653/v1/P19-1291. URL
https://www.aclweb.org/anthology/P19-1291.

13
[32] Peter J. Liu*, Mohammad Saleh*, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser,
and Noam Shazeer. Generating wikipedia by summarizing long sequences. In International
Conference on Learning Representations, 2018. URL https://openreview.net/forum?
id=Hyg0vbWC-.
[33] Yury A. Malkov and D. A. Yashunin. Efficient and robust approximate nearest neighbor search
using hierarchical navigable small world graphs. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 42:824–836, 2016. URL https://arxiv.org/abs/1603.09320.
[34] Gary Marcus. The next decade in ai: four steps towards robust artificial intelligence. arXiv
preprint arXiv:2002.06177, 2020. URL https://arxiv.org/abs/2002.06177.
[35] Luca Massarelli, Fabio Petroni, Aleksandra Piktus, Myle Ott, Tim Rocktäschel, Vassilis
Plachouras, Fabrizio Silvestri, and Sebastian Riedel. How decoding strategies affect the
verifiability of generated text. arXiv preprint arXiv:1911.03587, 2019. URL https:
//arxiv.org/abs/1911.03587.
[36] Nikita Moghe, Siddhartha Arora, Suman Banerjee, and Mitesh M. Khapra. Towards exploit-
ing background knowledge for building conversation systems. In Proceedings of the 2018
Conference on Empirical Methods in Natural Language Processing, pages 2322–2332, Brus-
sels, Belgium, October-November 2018. Association for Computational Linguistics. doi:
10.18653/v1/D18-1255. URL https://www.aclweb.org/anthology/D18-1255.
[37] Preksha Nema and Mitesh M. Khapra. Towards a better metric for evaluating question generation
systems. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language
Processing, pages 3950–3959, Brussels, Belgium, October-November 2018. Association for
Computational Linguistics. doi: 10.18653/v1/D18-1429. URL https://www.aclweb.org/
anthology/D18-1429.
[38] Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder,
and Li Deng. MS MARCO: A human generated machine reading comprehension dataset. In
Tarek Richard Besold, Antoine Bordes, Artur S. d’Avila Garcez, and Greg Wayne, editors,
Proceedings of the Workshop on Cognitive Computation: Integrating neural and symbolic
approaches 2016 co-located with the 30th Annual Conference on Neural Information Processing
Systems (NIPS 2016), Barcelona, Spain, December 9, 2016, volume 1773 of CEUR Workshop
Proceedings. CEUR-WS.org, 2016. URL http://ceur-ws.org/Vol-1773/CoCoNIPS_
2016_paper9.pdf.
[39] Rodrigo Nogueira and Kyunghyun Cho. Passage re-ranking with BERT. arXiv preprint
arXiv:1901.04085, 2019. URL https://arxiv.org/abs/1901.04085.
[40] Ethan Perez, Siddharth Karamcheti, Rob Fergus, Jason Weston, Douwe Kiela, and Kyunghyun
Cho. Finding generalizable evidence by learning to convince q&a models. In Proceedings
of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th
International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages
2402–2411, Hong Kong, China, November 2019. Association for Computational Linguistics.
doi: 10.18653/v1/D19-1244. URL https://www.aclweb.org/anthology/D19-1244.
[41] Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu,
and Alexander Miller. Language models as knowledge bases? In Proceedings of the 2019
Conference on Empirical Methods in Natural Language Processing and the 9th International
Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463–2473, Hong
Kong, China, November 2019. Association for Computational Linguistics. doi: 10.18653/v1/
D19-1250. URL https://www.aclweb.org/anthology/D19-1250.
[42] Fabio Petroni, Patrick Lewis, Aleksandra Piktus, Tim Rocktäschel, Yuxiang Wu, Alexander H.
Miller, and Sebastian Riedel. How context affects language models’ factual predictions. In
Automated Knowledge Base Construction, 2020. URL https://openreview.net/forum?
id=025X0zPfn.
[43] Alec Radford. Improving Language Understanding by Generative Pre-Training, 2018.
URL https://s3-us-west-2.amazonaws.com/openai-assets/research-covers/
language-unsupervised/language_understanding_paper.pdf.

14
[44] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya
Sutskever. Language models are unsupervised multitask learners, 2019. URL
https://d4mucfpksywv.cloudfront.net/better-language-models/language_
models_are_unsupervised_multitask_learners.pdf.
[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena,
Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified
text-to-text transformer. arXiv e-prints, 2019. URL https://arxiv.org/abs/1910.10683.
[46] Adam Roberts, Colin Raffel, and Noam Shazeer. How much knowledge can you pack into
the parameters of a language model? arXiv e-prints, 2020. URL https://arxiv.org/abs/
2002.08910.
[47] Stephen Robertson and Hugo Zaragoza. The probabilistic relevance framework: Bm25 and
beyond. Found. Trends Inf. Retr., 3(4):333–389, April 2009. ISSN 1554-0669. doi: 10.1561/
1500000019. URL https://doi.org/10.1561/1500000019.
[48] Sainbayar Sukhbaatar, arthur szlam, Jason Weston, and Rob Fergus. End-to-end memory net-
works. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances
in Neural Information Processing Systems 28, pages 2440–2448. Curran Associates, Inc., 2015.
URL http://papers.nips.cc/paper/5846-end-to-end-memory-networks.pdf.
[49] Sainbayar Sukhbaatar, arthur szlam, Jason Weston, and Rob Fergus. End-To-End Memory Net-
works. In C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, editors, Advances
in Neural Information Processing Systems 28, pages 2440–2448. Curran Associates, Inc., 2015.
URL http://papers.nips.cc/paper/5846-end-to-end-memory-networks.pdf.
[50] James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. FEVER: a
large-scale dataset for fact extraction and VERification. In Proceedings of the 2018 Conference
of the North American Chapter of the Association for Computational Linguistics: Human
Language Technologies, Volume 1 (Long Papers), pages 809–819, New Orleans, Louisiana,
June 2018. Association for Computational Linguistics. doi: 10.18653/v1/N18-1074. URL
https://www.aclweb.org/anthology/N18-1074.
[51] James H. Thorne and Andreas Vlachos. Avoiding catastrophic forgetting in mitigating model
biases in sentence-pair classification with elastic weight consolidation. ArXiv, abs/2004.14366,
2020. URL https://arxiv.org/abs/2004.14366.
[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,
Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. V. Luxburg,
S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural
Information Processing Systems 30, pages 5998–6008. Curran Associates, Inc., 2017. URL
http://papers.nips.cc/paper/7181-attention-is-all-you-need.pdf.
[53] Ashwin Vijayakumar, Michael Cogswell, Ramprasaath Selvaraju, Qing Sun, Stefan Lee,
David Crandall, and Dhruv Batra. Diverse beam search for improved description of com-
plex scenes. 2018. URL https://www.aaai.org/ocs/index.php/AAAI/AAAI18/paper/
view/17329.
[54] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman.
GLUE: A multi-task benchmark and analysis platform for natural language understanding.
In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting
Neural Networks for NLP, pages 353–355, Brussels, Belgium, November 2018. Association for
Computational Linguistics. doi: 10.18653/v1/W18-5446. URL https://www.aclweb.org/
anthology/W18-5446.
[55] Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix
Hill, Omer Levy, and Samuel Bowman. SuperGLUE: A Stickier Benchmark for General-
Purpose Language Understanding Systems. In H. Wallach, H. Larochelle, A. Beygelzimer,
F. d\textquotesingle Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information
Processing Systems 32, pages 3261–3275. Curran Associates, Inc., 2019. URL https://
arxiv.org/abs/1905.00537.

15
[56] Shuohang Wang, Mo Yu, Xiaoxiao Guo, Zhiguo Wang, Tim Klinger, Wei Zhang, Shiyu Chang,
Gerry Tesauro, Bowen Zhou, and Jing Jiang. R3 : Reinforced ranker-reader for open-domain
question answering. In Sheila A. McIlraith and Kilian Q. Weinberger, editors, Proceedings of
the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative
Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational
Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7,
2018, pages 5981–5988. AAAI Press, 2018. URL https://www.aaai.org/ocs/index.
php/AAAI/AAAI18/paper/view/16712.
[57] Shuohang Wang, Mo Yu, Jing Jiang, Wei Zhang, Xiaoxiao Guo, Shiyu Chang, Zhiguo Wang,
Tim Klinger, Gerald Tesauro, and Murray Campbell. Evidence aggregation for answer re-
ranking in open-domain question answering. In ICLR, 2018. URL https://openreview.
net/forum?id=rJl3yM-Ab.
[58] Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. In Yoshua Bengio
and Yann LeCun, editors, 3rd International Conference on Learning Representations, ICLR
2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. URL
http://arxiv.org/abs/1410.3916.
[59] Jason Weston, Emily Dinan, and Alexander Miller. Retrieve and refine: Improved sequence
generation models for dialogue. In Proceedings of the 2018 EMNLP Workshop SCAI: The 2nd
International Workshop on Search-Oriented Conversational AI, pages 87–92, Brussels, Belgium,
October 2018. Association for Computational Linguistics. doi: 10.18653/v1/W18-5713. URL
https://www.aclweb.org/anthology/W18-5713.
[60] Shiyue Zhang and Mohit Bansal. Addressing semantic drift in question generation for semi-
supervised question answering. In Proceedings of the 2019 Conference on Empirical Meth-
ods in Natural Language Processing and the 9th International Joint Conference on Natural
Language Processing (EMNLP-IJCNLP), pages 2495–2509, Hong Kong, China, Novem-
ber 2019. Association for Computational Linguistics. doi: 10.18653/v1/D19-1253. URL
https://www.aclweb.org/anthology/D19-1253.
[61] Wanjun Zhong, Jingjing Xu, Duyu Tang, Zenan Xu, Nan Duan, Ming Zhou, Jiahai Wang, and
Jian Yin. Reasoning over semantic-level graph for fact checking. ArXiv, abs/1909.03745, 2019.
URL https://arxiv.org/abs/1909.03745.

16
Figure 4: Annotation interface for human evaluation of factuality. A pop-out for detailed instructions
and a worked example appear when clicking "view tool guide".

A Human evaluation

Figure 4 shows the user interface for human evaluation. To avoid any biases for screen position,
which model corresponded to sentence A and sentence B was randomly selected for each example.
Annotators were encouraged to research the topic using the internet, and were given detailed instruc-
tions and worked examples in a full instructions tab. We included some gold sentences in order to
assess the accuracy of the annotators. Two annotators did not perform well on these examples and
their annotations were removed from the results.

B Further details on Open-Domain QA

For open-domain QA, multiple answer annotations are often available for a given question. These
answer annotations are exploited by extractive models during training as typically all the answer
annotations are used to find matches within documents when preparing training data. For RAG, we
also make use of multiple annotation examples for Natural Questions and WebQuestions by training
the model with each (q, a) pair separately, leading to a small increase in accuracy. For TriviaQA,
there are often many valid answers to a given question, some of which are not suitable training targets,
such as emoji or spelling variants. For TriviaQA, we filter out answer candidates if they do not occur
in top 1000 documents for the query.

TriviaQA Evaluation setups The open-domain QA community customarily uses public develop-
ment datasets as test datasets, as test data for QA datasets is often restricted and dedicated to reading
compehension purposes. We report our results using the datasets splits used in DPR [22], which are
consistent with common practice in Open-domain QA. For TriviaQA, this test dataset is the public
TriviaQA Web Development split. Roberts et al. [46] used the TriviaQA official Wikipedia test set
instead. Févry et al. [14] follow this convention in order to compare with Roberts et al. [46] (See
appendix of [14]). We report results on both test sets to enable fair comparison to both approaches.
We find that our performance is much higher using the official Wiki test set, rather than the more
conventional open-domain test set, which we attribute to the official Wiki test set questions being
simpler to answer from Wikipedia.

C Further details on FEVER

For FEVER classification, we follow the practice from [28], and first re-generate the claim, and
then classify using the representation of the final hidden state, before finally marginalizing across
documents to obtain the class probabilities. The FEVER task traditionally has two sub-tasks. The
first is to classify the claim as either "Supported", "Refuted" or "Not Enough Info", which is the task
we explore in the main paper. FEVER’s other sub-task involves extracting sentences from Wikipedia

17
as evidence supporting the classification prediction. As FEVER uses a different Wikipedia dump to
us, directly tackling this task is not straightforward. We hope to address this in future work.

D "Null document" Probabilities

We experimented with adding "Null document" mechanism to RAG, similar to REALM [18] in order
to model cases where no useful information could be retrieved for a given input. Here, if k documents
were retrieved, we would additionally "retrieve" an empty document and predict a logit for the null
document, before marginalizing over k + 1 predictions. We explored modelling this null document
logit by learning (i) a document embedding for the null document, (ii) a static learnt bias term, or
(iii) a neural network to predict the logit. We did not find that these improved performance, so in
the interests of simplicity, we omit them. For Open MS-MARCO, where useful retrieved documents
cannot always be retrieved, we observe that the model learns to always retrieve a particular set of
documents for questions that are less likely to benefit from retrieval, suggesting that null document
mechanisms may not be necessary for RAG.

E Parameters
Our RAG models contain the trainable parameters for the BERT-base query and document encoder of
DPR, with 110M parameters each (although we do not train the document encoder ourselves) and
406M trainable parameters from BART-large, 406M parameters, making a total of 626M trainable
parameters. The best performing "closed-book" (parametric only) open-domain QA model is T5-11B
with 11 Billion trainable parameters. The T5 model with the closest number of parameters to our
models is T5-large (770M parameters), which achieves a score of 28.9 EM on Natural Questions [46],
substantially below the 44.5 that RAG-Sequence achieves, indicating that hybrid parametric/non-
parametric models require far fewer trainable parameters for strong open-domain QA performance.
The non-parametric memory index does not consist of trainable parameters, but does consists of 21M
728 dimensional vectors, consisting of 15.3B values.

F Retrieval Collapse
In preliminary experiments, we observed that for some tasks such as story generation [11], the
retrieval component would “collapse” and learn to retrieve the same documents regardless of the
input. In these cases, once retrieval had collapsed, the generator would learn to ignore the documents,
and the RAG model would perform equivalently to BART. The collapse could be due to a less-explicit
requirement for factual knowledge in some tasks, or the longer target sequences, which could result
in less informative gradients for the retriever. Perez et al. [40] also found spurious retrieval results
when optimizing a retrieval component in order to improve performance on downstream tasks.

18