q124/random_k4/random_4.md

Encyclopedic VQA: Visualquestionsaboutdetailedpropertiesoffine-grainedcategories ThomasMensink†,∗ JasperUijlings†,∗ LluisCastrejon∗ ArushiGoel‡ mensink@google.com jrru@google.com lluisc@google.com goel.arushi@gmail.com FelipeCadar‡ HowardZhou∗ FeiSha∗ Andre´ Araujo∗ cadar@dcc.ufmg.br howardzhou@google.com fsha@google.com andrearaujo@google.com VittorioFerrari∗ vittoferrari@google.com Abstract We propose Encyclopedic-VQA, a large scale visual question answering (VQA) dataset featuring visual ques- tions about detailed properties of fine-grained categories and instances. It contains 221k unique question+answer pairs each matched with (up to) 5 images, resulting in a total of 1M VQA samples. Moreover, our dataset comes with a controlled knowledge base derived from Wikipedia, marking the evidence to support each answer. Figure1:Oneofthetwoanswersaboveiswrong,doyouknow Empirically, we show that our dataset poses a hard whichone? Encyclopedicquestionsaboutdetailedpropertiesof challenge for large vision+language models as they per- fine-grainedentitiesaredifficult.Notonlyforhumans,butalsofor form poorly on our dataset: PaLI [14] is state-of-the-art largeVLMs.PaLI[14]failstoanswerbothquestionscorrectly. on OK-VQA [37], yet it only achieves 13.0% accuracy on our dataset. Moreover, we experimentally show Answeringthesequestionscorrectlyrequiresknowledgeof that progress on answering our encyclopedic questions detailed properties (i.e. symbol of which city, year of au- can be achieved by augmenting large models with a tomation) of fine-grained categories (‘Pinus Pinea’) or in- mechanism that retrieves relevant information from the stances (‘Point Reyes Lighthouse’). We hypothesize that knowledge base. An oracle experiment with perfect re- this type of encyclopedic knowledge is hard for VLMs to trievalachieves87.0%accuracyonthesingle-hopportion properlyencodeinitsmodelparametersbecausesuchlong- of our dataset, and an automatic retrieval-augmented tail information occurs rarely in its training data (i.e. the prototypeyields48.8%.Webelievethatourdatasetenables web). Additionally,VLMsproducesuchincorrectanswers future research on retrieval-augmented vision+language generally with high confidence, while these answers are models. It is available at https://github.com/ hardforuserstoverifysincethemodeldoesnotprovideany google-research/google-research/tree/ explanations. Theanswerto Fig.1(left) iscorrect; (right) master/encyclopedic_vqa. shouldbe1975. PaLI[14]predictsbothincorrectly. Bothproblemscanbeaddressedbyretrieval-augmented 1.Introduction models, which base their predictions on knowledge re- trieved from a database. These models recently gained Recently, large Vision+Language models (VLMs) have popularity in NLP (e.g. [9, 23, 30, 33]), and some early demonstrated impressive performance on Visual Question multi-modal models also exist [22, 26, 36, 52]. Retrieval- Answering (VQA) benchmarks [7, 14, 24, 57]. However, augmentedmodelsarewell-suitedforencyclopedicknowl- Fig.1showstwotypicalexampleswheresuchmodelsfail. edge since retrieving the correct entry from a knowledge †Equalcontribution.‡WorkdoneduringinternshipatGoogle. base greatly facilitates constructing the right answer. Fur- ∗GoogleResearch thermore, the retrieved piece of knowledge provides attri- 3202 luJ 42 ]VC.sc[ 2v42290.6032:viXra

Question Type skramdnaL Templated Automatic Automatic - multi-answer 2-Hop Q: Who founded this monastery? Q: When was the first permanent Q: What fish can be found in this lake? Q: What amusement park is located in settlement made at this valley? the city where this square is located? A: Prince Constantin Brâncoveanu A: 1864 A: trout, lake char A: Tivoli Gardens C: Horezu monastery C: Clover valley C: Úlfljótsvatn C: Rådhuspladsen, Copenhagen dlroW larutaN Q: How old does this reptile become? Q: How many feet tall does this tree Q: Where is this bird found? Q: How many national park service maintained grow to? sites are in the state where this plant grows? A: 40 years A: 7 to 13 A: Colombia, Venezuela, Ecuador A: 24 C: Gila monster C: Acacia paradoxa C: Boissonneaua C: Chorizanthe rigida Figure2: ExampleVQAannotationsfordifferentquestiontypes. EachexampleconsistsofanimageI,aquestionQandtheanswer A.WealsoshowthecategoryCofthesubjectofthequestion.Asattribution,weprovideasectionwithintheWikipediapageofCwhich supportstheanswer.OurEncyclopedic-VQAdatasethasatotalof1M(I,Q,A)triplets. OK-VQA A-OKVQA FVQA KVQA S3VQA Q: What sort of vehicle uses Q: What does the man who Q: What furniture in this image Q: Who is to the left of Hillary Q: What is the process by this item? sits have trouble doing? can I lie on? Clinton? which this insect develops? A: firetruck A: walking A: sofa A: Aamir Khan A: complete metamorphosis Q: When was this piece of Q: What could block the Q: Which transportation way in Q: Who among the people in the Q: Where was this weapon used sporting equipment invented? washer’s door? this image is cheaper than taxi? image is the eldest? to attack the French camp? A: 1926 A: stove A: bus A: Person in the left A: Crecy Figure3: TypicalexamplesofVQAdatasetswhichrequireknowledgebeyondtheimage. bution to the answer by design. This increases model in- quiringknowledgeofdetailedproperties,butmostlyabout terpretabilityandthereforehumantrust,hasapplicationsto coarserbasic-levelcategories45. fairnessandhelpsdiagnosinganddebuggingmodelerrors. Thereareafewotherdatasetswhichtargetdetailedproper- To drive progress on handling encyclopedic knowledge ties[27,47]. Butboth[27,47]lackattributionannotations, and attribution in VQA we need a suitable dataset. The [27]isverysmall(Tab.1right-mostcolumn),and[47]fo- popular OK-VQA [37] and A-OKVQA [46] datasets fo- cusesonlyoncelebrities(Fig.3fourthcolumn). Henceno cus on questions requiring knowledge outside the query currentVQAdatasetisfullysatisfactory. image. However, the majority of questions require com- In this paper we introduce the Encyclopedic-VQA monsense knowledge (e.g. Fig. 3 top-row, two left-most dataset, which offers several attractive features (Fig. 2, examples), a type of knowledge for which current VLMs Tab.1). Ourdatasetasksquestionsaboutfine-grainedcat- are powerful in both theory and practice [7, 14, 24, 57]. egories from iNaturalist 2021 [51] (Fig. 2 bottom row) These datasets [37, 46] also include some questions re- andinstancesfromtheGoogleLandmarksDatasetv2[56]

OK-VQA[37] A-OKVQA[46] FVQA[53] S3VQA[27] KVQA[47] Encyclopedic-VQA(ours) Trulymultimodal

				++		++	++	++		++		++
Encyclopedic				±		-	-	+		++		++
(cid:44)→Detailedproperties				+		±	-	++		++		++
(cid:44)→fine-grainedcategories/instances				-		-	-	+		++		++
Controlledknowledgebase				--		--	+	±		±		++
(cid:44)→answersupportedbyKB				--		--	++	++		++		++
(cid:44)→KBprovided
				--		--	++	--		++		++
--------------------	---	---	---	---	---	---	---	---	---	---	---	---	---
(cid:44)→KBfree-form				--		--	--	--		--		++
(cid:44)→attribution				--		±	++	--		--		++
Scale				±		+	--	--		++		++
Two-hop				--		--	--	--		++		+
Subject various various various various celebrities fine-grainedspecies,landmarks
NumberoftextquestionsQ				14k		25k	6k	7k		183k		221k
--------------------------------	---	---	---	---	---	---	---	---	---	----	---	------	---
NumberofimagesI				14k		24k	2k	7k		25k		514k
NumberofuniqueVQAtriplets(I,Q,A)				14k		25k	6k	7k		183k		1,036k
Table1: ComparisonofrecentVQAdatasets.WecomparerecentVQAdatasetswithourproposeddatasetonVQAdesignprinciples.
(Fig. 2 top row). We construct questions about detailed that specific part of the knowledge base which supports it.
properties based on Wikipedia (e.g. founder of building, Hencetheknowledgebaseisanintegralpartofourdataset,
maximumageofanimal). Asaconsequence,allquestions whichenablesmeasuringwhetheramodelanswersaques-
in our dataset are about encyclopedic knowledge. Further- tion correctly for the right reason. For generality we want
more, we provide a controlled knowledge base suited to thisknowledgebasetobefree-formtextandtocontainim-
answerthesequestions: 2MWikipediapagesconsistingof ages. (4) Scale. The dataset should be large. Dataset size
free-formtextandimages[48]. Wealsomarkground-truth hasalwaysmatteredandthisisevenmoretruewiththein-
attributionforeachansweratthegranularitylevelofasec- creasinglylargeVLMmodels(e.g.[6,14,28,60,39]). (5)
tion within a Wikipedia page. Importantly, our dataset is Two-hop. Aportionofourquestionsshouldrequireknowl-
collected at scale: we have 221k unique question+answer edge from multiple different documents from the knowl-
pairseachmatchedwitharound5images,resultinginato- edgebase. Includingsuchcomplextwo-hopquestions[58]
talof1Mexamples.Thismakesourdatasetthelargestofits leavessubstantialheadroomforfuturemodeldevelopment.
kind. Finally,	manyof	ourquestions		arecomplex		two-hop
--------------	------	------------	---	----------	---	-------	---	---	---	---	---	---	---
3.RelatedWork
questions[58],whichrequiremultipledifferentdocuments
fromtheknowledgebasetosolve(Fig.2right).
VisualQuestionAnswering(VQA).DAQUAR[35],FM-
We validate	the usefulness		of our	dataset	through	sev-
-----------	--------------	---	------	-------	-------	----	---	---	---	---	---	---	---
IQA[20],VisualMadlibs[59],VQAv1[8]andVQAv2[21]
eralexperiments.	Inparticular,wedemonstratethatalarge
----------------	------------------------------------	------	----------------	-----------	------	----------	---------	--------	-----------	--------------	---	----------	------
							are early	VQA	datasets.	These datasets		mostly ask	visual
VLM (PaLI	[14]) which	yields	state-of-the-art			results on
							questions	that can	be answered	based	on	the query	image
OK-VQA	[37] performs	poorly	on	our dataset	(13.0%	ac-
andgenericknowledgesuchasbasic-levelcategoryrecog-
curacy).	Next we demonstrate		through		an oracle	experi-
--------	-------------------	---	-------	---	---------	-------	---	---	---	---	---	---	---
nition(e.g.‘cat’),countingitems,colors,etc.
mentthatretrieval-augmentedmodelscanyield87.0%ac-
							Knowledge-based		VQA.	Tab.	1 and	Fig. 3 compare
-------	---------------	---	------	-----	------	---------	---------------	-----	------------	----	-------	--------------	---
curacy.	Finally, we use	an	online	image	search	engine to
							datasets	which	are designed	to	require	knowledge	not
buildanautomaticretrieval-augmentedprototype,reaching
							presentintheimage[27,37,46,47,53].					Inparticular,OK-
----------------	-----------------------	--------	-----	--------	-----------	---------	----------------------------------	----------	---------------------------------	----	--------	----------------	---
48.8% accuracy.	We	conclude	that	(1) our	dataset	poses a
							VQA [37]	and	A-OKVQA	[46]	mostly	require common-
strong challenge	in VQA	and	is in	fact too	hard	for stan-
							senseknowledge.		Inaddition,somequestions(18%)inA-
dard VLMs;	(2) retrieval-augmented			VLMs	demonstrate	a
							OKVQA	do require	knowledge	of	detailed	properties,	but
strongpotentialforaddressingencyclopedicknowledge;(3)
							about basic-level		categories.	Finally,	3%	of the questions
---	---	---	---	---	---	---	-----------------	---	-----------	--------	---	----------------	---
ourresultsleavesignificantheadroomforfurtherresearchto
							requireknowledgeaboutphysics.				Inthispaperwecreatea
---	---	---	---	---	---	---	-----------------------------	---	---	---	--------------------	---	---
improveretrieval-augmentedVLMs.
datasetwithquestionsexclusivelyaboutdetailedproperties
							of fine-grained		categories	and instances		(Fig. 2).	We be-
---	---	---	---	---	---	---	---------------	---	----------	-------------	---	---------	------
2.Designprinciplesforourdataset
							lieve that	such encyclopedic		questions	truly	require	access
---	---	---	---	---	---	---	----------	-----------------	---	---------	-----	-------	------
We create our VQA dataset with the following desired toaknowledgebasetobeanswered. Infact,wereleasethe
properties in mind: (1) Truly Multimodal. The questions knowledgebasealongwithourdataset,whereasnoexplicit
should not be answerable without the image or the textual knowledgebasewasinvolvedinthecreationof[37,46].
question [21]. (2) Encyclopedic. The questions should SomeexistingVQAdatasetsaresupportedbyaknowl-
be about detailed properties of fine-grained categories or edgebase[27,47,53]. FVQA[53]isaboutcommonsense
instances; a type of questions which are problematic for knowledge. KVQA [47] asks detailed properties about
vanilla VLMs [7, 14, 24, 57]. (3) Controlled knowledge celebrities. Both[47,53]arebasedonastructuredknowl-
base. Each answer in our dataset should be attributable to edge base (RDF triplets). Instead, our knowledge base is

readilyavailablefree-formWikipediatext,whichhasmore								4.Dataset
information.	Furthermore,KVQAisexclusivelyaboutpeo-
Wenowdetailtheconstructionofourdatasetwhilefol-
ple, which	requires	specialized		methods	based	on	face de-
----------	----------------	-----------	------------	-------	-----	-------	--------	-----------	------------	----------	--------	---------	------------	-------
								lowing the	design	principles		(Sec. 2).	To achieve	scale
tection	and recognition.		In contrast,		our	dataset	offers a
								we automate	construction		whenever		possible. We	use hu-
broaderrangeoftopics,includinganimals,plantsandland-
manannotatorstoensurequalityandtoprovideinformation
marks.	S3VQA	[27]	is the most	related	work	to	ours. It
------	-----	----	-----------	-------	----	---	--------	--------------------------------	---	---	---	---	---------------	---
								whichwecouldnotgetautomatically.					Wesimplifyhuman
buildsoncategoriesofOpenImages[32],somefine-grained
annotationtasksasmuchaspossible,whichincreasesboth
and others	more	basic-level.		They automatically		generate
----------	----	------------	---	------------------	---	--------	---	---	---	---	---	---	---	---
thequalityandefficiencyoftheirwork.
questions	from	Wikipedia	articles,		which	by construction
---------	----	---------	---------	---	-----	---------------	---	---	---	---	---	---	---	---
WedefineaunitforourVQAtaskasatriplet(I,Q,A).
are about	detailed	properties.		Then	they let	experts	manu-
---------------------------------------	---------	-----------	--------	----	----------	------------	-----	--------------------------------------	---	---	---	------------------------	---------	---
								Themulti-modalquestionfeaturesanimageI					andaccom-
allyselectandrephraserelevantquestions.						Ourpaperalso
								panyingtextualquestionQ.				ThesubjectofthequestionQ
automatically	generates		QA pairs	from	Wikipedia.		How-
appearsinI
									andisacoreconceptofourwork.				Wereferto
---------------------------------	---	---	---	---	---------------	---	---	---	---------------------------	---	---	---	---------	---
ever,wegobeyond[27]inseveralways:					(1)ourdatasetis
thecategoryofthesubjectasC(e.g.‘HorezuMonastery’or
muchlarger(7kvs1MVQAtriplets),(2)weincludemore
								‘GilaMonster’inFig.			2-left).	TheanswerAispurelytex-
---------	---------	----------	------	------	-------	----	--------	---------------------------------------------------	---	-----	--------	----------------------	----------	---------
complex	multi-hop	questions,		(3) we	release	the	knowl-
								tual. Ourdatasetalsorecordstheevidenceforeachanswer
edge base	along	with	the QA	pairs.	This	also	includes
								in the knowledge		base,	as a	section	within the	Wikipedia
ground-truthanswerattributionintheformofthesupport-
								pageforC	wheretheanswerisfound.				Incontrasttomany
---	---	---	---	---	---	---	---	--------	----------------------	---	---	---	----------------	---
ingWikipediasection.
								existing	VQA	datasets,	we	record only	a single	answer,
---	--------	-------	----	-------------	---	------------	---	--------	---	---------	---	-----------	--------	-------
The	InfoSeek	dataset	[15]	is concurrent		to our work.	It
sinceitisunambiguousgiventheknowledgebase.
also targets	detailed	properties		of fine-grained		categories,
------------	-----------	------------	---	---------------	-------	-----------	---	-------	---	-------	-----	------------	---	----------
								To make	our	dataset	truly	multi-modal,	we	always re-
but does	not include	multi-answer		and	two-hop	questions,
fertothesubjectofthequestionbyitssupercategory(e.g.
andwascollectedusingadifferentannotationprotocol.
								‘this monastery’,		‘this	reptile’,	as opposed	to ‘the	Horezu
----------------------------	---	---	---	------------------	---	---	---	-----------------	----	-----	---------	----------	------------	------
Automaticquestiongeneration.				AfewVQAdatasetsare
								Monastery’,	‘the	Gila	monster’,	etc.).	This ensures	visual
generated from image captions [44, 13]. While this en- recognitionisrequiredtosolveourtask,astherewouldbe
ablesmuchlargerscale,itresultsinsimplevisualquestions.
manypotentialanswerstothetextualquestionalone(e.g.a
We build	a portion	of	our dataset	using	a	similar	pipeline
--------	---------	---	-----------	-----	---	-------	--------	---	---	---	---	---	---	---
differentanswerforeachmonasteryintheworld).
as [13],	but apply	it to	Wikipedia	pages	instead	to	obtain
--------	---------	-----	---------	-----	-------	---	------	---	---	---	---	---	---	---
4.1.SupportingDatasets
high-quality	questions		on detailed	properties.		As	another
------------	---------	---	-----------	-----------	---	---	-------	---	---	---	---	---	---	---
difference,weverifythemwithhumanannotators. We build on two of the largest existing datasets with
Multi-Hop	Reasoning		in NLP.	The	NLP	community	has
---------	---------	---	-------	---	---	---------	---	-----------------------	---	---	------------------------------	---	---	---
								fine-grainedcategories:			iNaturalist2021(iNat21)[51]and
many text-only QA datasets (e.g. [19, 29, 31, 38, 42, 50, Google Landmarks Dataset V2 (GLDv2) [56]. iNat21 is
55, 58]). Notably, HotpotQA [58] introduces the concept afine-grainedvisualclassificationdatasetcontaining2.7M
of‘bridgeentity’whichlinkstworelatedentities.						Theyuse
---------------------------------------------	---	---	---	---	---	-------	---	---	---	---	---	---	---	---
imagesdepicting10,000species,groupedinto11supercat-
thistoshowanannotatortworelatedWikipediaparagraphs egories: plants, insects, birds, mammals, etc. GLDv2 is
from different	pages	and	ask	to create	questions	which	re-
--------------	-----	---	---	---------	---------	-----	---	----------	-----------	---	-------	----	---------	---------
								a landmark	recognition		dataset	with	4M images	depicting
quires knowledge from both paragraphs. We use ‘bridge 200k landmarks. Each landmark is sourced from Wiki-
entities’toautomaticallychaintwoquestionstogetherinto media Commons [5], enabling us to mine the existing
acompoundtwo-hopquestion.
category-hierarchy(allowingtodefinesupercategories,like
Retrieval-augmented models. Our dataset seems partic- bridges, castles, lighthouses, lakes etc.), and to link them
ularly suited for retrieval-augmented models. While pio- to Wikipedia articles. We use these provided annotations
neered in NLP [9, 23, 30, 33], several recent works use to speed up the annotation of our dataset, to identify rele-
retrieval for VQA. KRISP [36] leverages triplets encoding vantknowledgecategories,andtoassignimagesItotextual
facts, categorical knowledge and object relationships in a questionsQautomatically.
graph-based reasoning framework. KAT [22] uses Wiki- We create our controlled knowledge base starting from
data triplets, and a reasoning module cross-attends them the WIT dataset [48], which contains 37M image-snippet
withGPT-3answers,theresultbeingfedintoadecoderfor pairs with 11M unique images from Wikipedia. We select
answer generation. InFactuality [52] leverages index im- allimageslinkedtoanEnglishsnippetandthenextendWIT
agestolinkentitiespresentinthequeryimageandWikidata to include the full corresponding Wikipedia article (snap-
triplets to gather facts, which are fed into a UNITER [16] shot of 13 August 2022). This spans 2M English articles.
reasoning module. REVEAL [26]’s external knowledge is We then identify which categories of iNat21 and GLDv2
composedofimage-textpairs,questionansweringpairsand map one-to-one to a single Wikipedia article. We found
knowledgegraphtriplets,whichareusedtoassistagenera- such unique mappings for 80% of the iNat21 and 50% of
tormoduleinproducinganswers. the GLDv2 categories. We only create questions for those

Monastery Horezu Monastery Templated question based upon super category Horezu Monastery Question: Who founded this monastery? The Horezu Monastery or Hurezi Monastery was founded in 1690 by Prince

			Constantin Brâncoveanu in the town of Horezu, Wallachia, Romania. It is					Answer:
considered to be a masterpiece of "Brâncovenesc style", known for its
			architectural purity and balance, the richness of its sculpted detail, its					Evidence:		Copy selected evidence
---	---	---	---------------------------------------------------------------------------	---	---	---	---	---------	---	----------------------	---	---
			treatment of religious compositions, its votive portraits, and its painted							Not enough evidence
decorative works.
Plant Acacia paradoxa Acacia paradoxa Automated question from wikipedia section
Are the following question/answer pairs relevant and correct?
Description
The large shrub or tree up to 2 to 4 metres (7 to 13 ft) tall and has a similar Question: How many feet tall does the acacia Yes
paradoxa grow to?
			width, it has ribbed branchlets that are often arched downward. It is dense					Answer:	7 to 13			No
---	---	---	----------------------------------------------------------------------------	---	---	---	---	--------	-------	---	---	---
with foliage; the leaves are actually enlarged petioles known as phyllodes.
They are crinkly and the new ones are covered in hairs. The erect phyllodes
								Question:	What are the new leaves of this plant			Yes
---	---	---	-------------------------------------------------------------------------	---	---	---	---	----------	--------------------------------------	---	---	---
			are asymettric and have a lanceolate shape and are around 30 millimetres						covered in?
(1.18 in) in length and 7 mm (0.276 in) wide. Answer: hairs No
Figure4: Datacollectionfortemplatedandautomaticallygeneratedsingle-hopquestions. Top: Expertscreatetemplatedquestions
Q based on a super category (e.g. Monastery). Annotators are given the full Wikipedia page for a particular C, and asked to provide
theanswersAandtheevidence. Bottom: QuestionsareautomaticallygeneratedfromaWikipediasectionofC andvalidatedbyhuman
annotators.Inbothprocesses,theannotatorneverseestheimageIcorrespondingtoC.
categorieswithauniquelyidentifiableWikipediaarticle. resolvethespecificcategoryC andamodelsimplyremem-
The combination of iNat21 & GLDv2 with Wikipedia beringQ+Apairscannotsolvethechallenge.
as knowledge base enables creating (I,Q,A) triplets cov- Automaticallygeneratedsingle-hop. Weincreasethedi-
ering the long-tail: questions about detailed properties (on versity of questions in the dataset with automatically gen-
Wikipedia	text)	of fine-grained	categories	(species	from
---------	-----	---------------	----------	--------	----	----------------	---	-----	-----------	--------	---	--------
						erated questions		(Fig.	4). Similar	to [27],	we	feed the
iNat21) and instances (landmarks from GLDv2). Hence, Wikipedia article of a category to a question generation
thissatisfiestheencyclopedicdesignprinciple. model [13]. It processes a section of an article at a time,
						producing	a large	number	of Q+A	pairs	(typically	100s),
-----------------------	---	---	---	---	---	---------------------------------	-------	------	------	----------------------	----------	------
4.2.Single-hopquestions						byinvertingstatementsfromthetext.
						Weincludetwofilteringsteps.				First,werequirethecat-
Theanswertoasingle-hopquestioncanbefoundinthe
						egory name	of	C to be	used in the	question.	This	reduces
------------------------------	---	---	-------------------	---	---	----------	---	-------	-----------	---------	----	-------
WikipediaarticleofthecategoryC			whichisthesubjectof
thequestion.Weconstructsuchquestionsintwoalternative thenumberofquestionsby20×,andensuresthattheques-
						tion is about	C	(and not	about another	entity	found	in the
---	---	---	---	---	---	-------------	---	--------	-------------	------	-----	------
ways: templatedandautomaticallygenerated.
						snippet	of text).	Second,	to increase	diversity	we	remove
--------------------	---	----------------------------	---	---	---	--------------	---------	---------	-----------	----------	------------	------
Templatedsingle-hop.		Weuseofthesupercategoryanno-
						near identical		questions	and limit	the number	of questions
tationsavailableinthesupportingdatasets(e.g.reptilesfor
perWikipediasection.
iNat21, monasteries		for GLDv2).	For each	super	category
-------------------	---	-----------	--------	-----	--------	---	---	---	---	---	---	---
wemanuallydefineseveralquestions,likeWhofoundedthis The human annotation task is now extremely simple:
validatewhethertheQ+Apairisrelevantandcorrect,given
monastery?.	Wethenaskhumanannotatorstoanswerthese
-----------	-------------------------------------	---	---	---	---	---	---	---	---	---	---	---
questionsforaparticularspecies/landmarkC (e.g.Horezu theWikipediasectionusedtogenerateit. AquestionQis
						relevant	if it is	is well-formed,	makes	sense	semantically,
---	---	---	---	---	---	--------	--------	---------------	-----	-----	-------------	---
Monastery),showingthemtherelevantWikipediapage(see
Fig. 4). Hence, theannotatorsdonothavetorecognizeC and is expected to have an unambiguous answer. The an-
in an image, which would require expert knowledge. In- swer A is correct if it is supported by the given Wikipedia
						section.	Finally,	wecandirectlyusetheWikipediasection
---	---	---	---	---	---	--------	--------	-----------------------------------	---	---	---	---
stead,theyjusthavetoidentifywhethertheanswerisinthe
Wikipediaarticle. Ifyes,theymarktheevidenceforit(for usedtogeneratetheQApairasevidencefortheanswer(at-
tributionforfree!)
attribution).	If	no, we discard	that question	(as	it cannot
-------------	---	--------------	-------------	---	---------	---	---	---	---	---	---	---
be answered with our knowledge base and it would break About 80% of the questions sent to annotators are val-
thecontrolledknowledgebaseprinciple). Westartwith10 idated positively. The questions contain (by design) the
questions per category, resulting in 4.3 questions on aver- name of the category C, and we rephrase them using a
ageperspecieswithavalidanswerforiNat21,and3.4per carefully-promptedFLAN[18]versionofPaLM[17]tore-
						place the	mention	of	C by its super	category,	which	pro-
---	---	---	---	---	---	---------	-------	---	--------------	---------	-----	----
landmarkforGLDv2(seeexamplesinFig.2firstcolumn).
E´glise
NowwehaveatextualquestionQandananswerA. To duces Q. For example, the question Where is the
form a complete multi-modal question, we pair Q with an Saint-Cannat located? is rephrased by PaLM to Where is
						thischurchlocated?		AnimageI		fromC
---	---	---	---	---	---	------------------	---	--------	---	-----	---	---
imageI ofC fromthesupportingdatasets. NotethatanyQ fromasupporting
occursmultipletimesinthedatasetwithdifferentanswers. datasetisthenassociatedtothisquestion.
HencebydesigntheaccompanyingimageI isnecessaryto Multi-answer questions. For many properties of a cat-

Q: What is the main competitor for food for this animal? Q: What is the population size of this animal? Chained two hop question with bridge: spotted hyena What is the population size of the main competitor for food of this animal? Q: What is the main competitor for food for this animal? Q: What is the population size of this animal? Chained question with bridge: spotted hyena

	Lion			Spotted hyena
Predatory competition
				The spotted hyena (Crocuta crocuta), also		What is the population
---	---	---	---	------------------------------------------	---	-----------------------	---
Lions and spotted hyenas occupy a similar known as the laughing hyena ... It is listed size of the main
as being of least concern by the IUCN on
ecological niche and compete for prey and a c c o u n t o f it s w id e s p r e a d ra n g e a n d la r g e competitor for food of
c a r ri o n ; a re v ie w o f d a t a a c ro ss s e v e ral n u m b e r s e s t im a te d b e t w e e n 2 7 , 0 0 0 a n d this animal?
	s tu d i e s i n d ic	a te s a d i e t a r y o ve rl a p o f
---	--------------------------------------	----------------------------------------------	---	---------------------------------------	---	---	---
	58.6%. Lions typically ignore hyenas…			47,000 individuals. The species is, ...
Figure 5: Illustrationoftwo-hopquestiongenerationusingbridgeentities. Thebridgeentity,e.g.spottedhyena,isananswertoa
single-hopquestionwhichhasitsownentryintheknowledgebase. Hencewecanaskasecondsingle-hopquestionaboutit. Thesetwo
questionsarethenchainedintoatwo-hopquestion.
egory C the response should be a list with multiple an- pair is combined with (up to) 5 images I showing differ-
swers. For example What fish can be found in this lake? ent instances of the same category C from the supporting
orWhereisthisbirdfound?,couldbeansweredwithalist datasets. Thisincreasesvisualdiversityandintotalweuse
offish(orcountries,respectively;Fig.2,thirdcolumn).For 514kuniqueimages. Thereareintotal15ktextualsingle-
these kinds of questions we define the multi-answer ques- hoptemplatedquestions,and158kautomaticallygenerated
tion type. We filter the automatically generated QA pairs questions. Moreover,thedatasetcontains25kmulti-answer
for questions likely to have multiple answers, and extract questionsand22ktwo-hopquestions.
aninitiallist.	Thenweasktheannotatorstocompleteand
--------------	-----------------------------------	---	---	---	---	---	---
Lookingindetailatthe221kQ+Apairs,175kofthem
verifythelistwithallthepossibleanswers.
haveauniqueQ(i.e.nostringmatchwithanyotherques-
4.3.Two-hopquestions tion). Moreover, there are 95k unique answers, of which
				73koccuronlyonce.	Theremaining22kanswersfollowa
---	---	---	---	-----------------	-----------------------------	---	---
Two-hopquestionsaredefinedasrequiringtworeason- Zipf’slaw:11kanswersoccurtwice,2kanswersoccur10+
ing steps	to obtain the	final answer.	Here we focus	on
---------	-------------	-------------	-------------	---	---	---	---
times,andlessthan400answersoccurmorethan50times.
chained two-hopquestions,whichrequiretwoconsecutive
retrievalsteps. Toconstructsuchquestionsatscale,weuse Train / Val / Test splits. To allow for evaluating differ-
thenotionofabridgeentity[58]: iftheanswertoasingle- entpropertiesofthedataset, wecarefullydesignitstrain/-
hop question is an entity with its own Wikipedia page, it val/test splits. First, there is no overlap in the images I
servesasabridgeentity,aboutwhichwecanaskasecond usedinthetrain,val,ortestsplits. ForGLDv2,wesample
single-hopquestion,seeFig.5. images for all our splits from their train-clean split, which
We automatically identify bridge entities from the avoids most noisy labels in their dataset. For iNat21, we
single-hopanswers(Sec.4.2),thenmanuallydiscardwords sampleourtrainimagesfromtheirtrainset,andourvaland
like yes, blue, and heavy which would yield very artificial test images from their validation set (their test set annota-
multi-hop questions. We then automatically generate and tionsarenotpubliclyavailable). Next, abouttwo-thirdsof
validate questions for these bridge entities using our auto- thequestionsQandhalfoftheanswersAinourvalandtest
maticallygeneratedsingle-hoppipeline. splits do not occur in our train split. Finally, roughly 17%
Wecreatethefinaltwo-hopquestionautomaticallyusing ofthesubjectcategoriesC inourvalandtestsplitsarenot
PaLM, by feeding it the two single-hop questions with an used in our train split. This allows for analyzing different
appropriateprompt.Unfortunately,thesetwo-hopquestions desirablepropertiesofVQAmodelsonourdataset.
areoftenincorrect.Therefore,wecreateasecondpromptto
validatethetwo-hopquestion:PaLMisaskedtoanswerthe
two-hopquestionusingthetwoinitialsingle-hopquestions
with answers	as context.	If the predicted	answer is	iden-
------------	-----------	----------------	---------	-----	------------------	--------------------	---
					Numberof(Q,A)pairs	Total(I,Q,A)triplets
tical to the answer of the second single-hop question, then Train Val Test Train Val Test
the two-hop question is validated and kept in our dataset; Templated 13,928 400 1,000 66,535 1,827 1,000
				Automatic	153,441 1,750	2,750 737,114	8,025 2,750
---------------------------------	----------------	-------------------	------------	-----------	-------------	-------------	-----------
otherwise,	it is discarded.	This increases	the accuracy	of
				MultiAnswer	23,929 400	1,000 112,736	1,844 1,000
thetwo-hopquestionssubstantially.		Examplesofvalidated
				TwoHop	21,040 400	1,000 99,866	1,895 1,000
two-hopquestionsaregiveninFig.2(right).
				Total	212,338 2,950	5,750 1,016,251	13,591 5,750
---	---	---	---	-----	-------------	---------------	------------
4.4.Datasetstatistics
Table2:
				DatasetStatistics.		Wereportthenumberof(question,
---	---	---	---	------------------	---	-----------------------------	---
Our dataset contains in total 1M (I,Q,A) triplets answer)pairs,and(image,question,answer)tripletsfordifferent
(Tab. 2). These are derived from a total of 221k textual question types. In total our dataset contains 1M VQA triplets,
makingitthelargestofitskind.
Q+Apairsfrom16.7kdifferentcategoriesC.			EachQ+A
--------------------------------------	---	---	-------	---	---	---	---

Dataset varietyofQ disparityofQ Model Retrieval OK-VQAAccuracy

		#uniquebigrams			avg.cosinedist.				KAT[22]		✓		53.1
								REVEAL[26]			✓		59.1
FVQA[53]				7.9k		0.620
								PromptCap[25]			-		60.4
KVQA[47]				12.8k		0.504
									PaLI[14]		-		64.5
OK-VQA[37]				19.3k		0.843
S3VQA[27]				19.7k		0.805
							Table	4:	OK-VQA:	PaLI	is state-of-the-art		on	OK-VQA.	It is
A-OKVQA[46]				33.9k		0.856
thereforeangoodcandidatetoapplytoourEncyclopedic-VQA.
Encyclopedic-VQA			257.9k			0.833
--------------------------------	------------	------------	-------------------------	--------	--------	--------	-------------	-----------	---	----------	----	----	---	-----	-------
Table	3: Diversity	of questions	Q	measured	in terms	of vari-	5.Experiments
etyanddisparity(higherisbetter).			Disparitynumbersaremostly
								We evaluate		PaLI [14],	PaLM	[17]	and	GPT-3	[10] on
takenfrom[46],butwereproducedtheirA-OKVQAresulttoval-
idate our re-implementation. Our dataset has the largest variety our dataset. We apply all models directly and in retrieval-
andisclosetothebestonesondisparity. augmented settings. For evaluation we use the test split
							of	our dataset		on the	single-hop	templated		and automat-
---	---	---	---	---	---	---	------	-----------	------------	---------	----------	-----------	-----------	------------	---
							ically	generated		questions	(except		for Sec.	5.6 where	we
							also	evaluate	multi-answer			and two-hop	questions).		We
4.5.QuestionDiversity
measureaccuracyasthepercentageofquestionswherethe
							predicted		model	answer	matches	the	ground-truth		answer.
---	---	---	---	---	---	---	---------	---	-----	------	-------	---	------------	---	-------
WewanttocomparethediversityofourVQAquestions
to other datasets. However, diversity is a broad concept All ground-truth answers and model predictions are pre-
							processed		following	standard		VQA	practices	[21]	to facil-
---	---	---	---	---	---	---	---------	---	---------	--------	---	---	---------	----	---------
whichmanifeststhroughacombinationofthreebasicprop-
							itate	checking	their	correctness		(remove	articles,	punctua-
-------	--------	----------	-----------	-----	----	--------	-----	--------	-----	-----------	---	-------	---------	--------	---
erties:	variety,	disparity,	and balance	[49].	Here	we focus
onvarietyanddisparity. tion, etc.). Furthermore, weusethe’BERTMatching’cri-
							terion	BEM	[12]	to determine		whether	a predicted		answer
---	---	---	---	---	---	---	------	---	----	------------	---	-------	-----------	---	------
Variety. This refers to the semantic variety spanned by is correct given the question and the ground-truth answer.
the VQA	questions.	One	simple	measure	of variety	is the
-------	----------	---	------	-------	----------	------	---	----------	---	------	---	----------------	---	------	------
							BEM	evaluation		allows	for	more flexibility		in the	answer
numberofuniquetextualquestionsQ(175kinourdataset).
							formulation		than	classical	exact	matching,		coming	much
------------------	----	-------------------------------	---------	----	---------	-------	-----------	---	-----	---------	-----	---------------	---	------	-----
However,	this	ignores the	fact that	some	questions	are se-
							closer	to	human	judgement		of correctness,		as is	shown
manticallysimilar.		Anothermeasureisthenumberoftop-
in[12]andinouruserstudyinAppendixB.Weconsideran
icscoveredbythequestions,forwhichwecanusethenum- answercorrectwhenitsBEMscoreis≥0.5. Ifanexample
ber of	categories	C (16.7k	in our	dataset).	However,	each
------	----------	--------	------	---------	--------	----	---	--------	------------	---	--------	---	----------	---	---
							has	multiple	ground-truth		answers,		we compute	the	BEM
datasettypicallyusesadifferentsetofcategoriesatdiffer-
scoreforeachandchoosethemaximum.
entlevelsofgranularity,makingtheircategorycountshard
to compare directly. So instead, we consider the number 5.1.LargeModelswithoutRetrieval
of unique	bigrams	across	all questions	Q,	as a	reasonable
---------	-------	------	-------------	---	----	----------	-----	---	-------	---	-----------	---	----	---------	---
							PaLI.	To	measure	the	performance		of a	large VLM	on
approximationofthesemanticspacespannedbyallQ.
Encyclopedic-VQA
										we	use	PaLI [14].	It yields	the	state-
----------	---	-------------	-------------	---	-------------	---	----------	---	--------	---	------	----------	---------	---	------
Disparity.		This measures	how different		the questions	are
							of-the-art		accuracy	on	OK-VQA	[37]	(64.5%),	and	out-
from each other. We follow the measure introduced in A- performs retrieval-augmented models such as KAT [22]
OKVQA	[46]:	(1) we	first project	all	questions	Q into
--------	-----	--------------	-------------	----------	---------	------	-------	---	----------	---	----	-------	---	-------	------
							(53.1%)		and REVEAL		[26]	(59.1%)	by	a large	margin
a common		semantic space	using	a publicly	available	[3]
(Tab.4). Thisdemonstratesthatthetypeofknowledgere-
Sentence-BERT model [43]; then (2) we measure the av- quired to solve OK-VQA can be captured by large VLMs.
eragecosinedistancebetweenallquestionpairs(definedas
							PaLI	is therefore		a good	candidate		to verify	whether	large
---------	----------	---	-----------	----	-----	---------	----	------------	---	------	---------	---	---------	-------	-----
1− cosine	similarity	by	[46]). Note	that	while	[46] pro-
VLMscapturealsoencyclopedicknowledge.WeuseaPaLI
posedthisasagenericmeasureofdiversity,itactuallyonly
modelwith17Bparameterspre-trainedonahugeamountof
measuresdisparityasdefinedin[49].
dataincludingWikipedia(soithasseentheknowledgebase
Results.							containingtheanswersforourdataset).						OurPaLImodelis
--------	---	--------------	---	---------	---	----------	-----------------------------------	---	---	---	---	---	--------------	---	---
	We	report variety	and	disparity	for	knowledge-
basedVQAdatasetsinTab.3. Ourdatasetoffersthelargest additionallyfine-tunedonOK-VQAandthereforeparticu-
questionvarietybyanorderofmagnitude.Importantly,this larlysuitedforVQAtasks. WefeedthismodelQ,I inputs
is not only due to dataset size: while Encyclopedic-VQA foreachtestsampletoproducemodelanswers.
andKVQA[47]arethelargestdatasetsandhaveacompa- Tab.5(firstrow)showsthatPaLIhasanaccuracyofonly
rable number of questions, KVQA has much less variety. 13.0%. This low performance indicates that PaLI either
In terms of disparity, our dataset is close to the best ones fails to recognise fine-grained categories and instances, or
OK-VQA[37]andA-OKVQA[46],andbetterthanFVQA, failstoprovidedetailedproperties,orboth. Henceencyclo-
KVQAandS3VQA.							pedicknowledgeistrulydifficultforthismodel.
-------------	---	---	---	---	---	---	-------------------------------------------	---	---	---	---	---	---	---	---

Model

	System		Retrieval	Extrainput		PaLI[14]	PaLM[17]	GPT-3[10]
	Vanillamodel		-		-	13.0%	19.7%	15.5%
			SubjectC		-	16.7%	31.0%	26.9%
	Oracle		KBArticle		-	29.7%	78.4%	77.4%
			KBSection		-	48.8%	87.0%	82.1%
			KBArticle		-	21.4%	48.0%	44.9%
Lens
			KBSection		-	28.1%	48.8%	44.6%
---	---	---	---------	---	---	-----	-----	-----	---
Captions
	PromptCap		-			17.8%	29.7%	25.6%
---	---------	---	---	---	---	-----	-----	-----	---
NNtrainsamples
Table 5: Accuracy	on single-hop	questions.
-----------------	-------------	----------	---	------------	--------	---	--------------	---------------------------	---------
			We	report model	accuracy	for	our single-hop	templated and automatically	generated
questions. WhilelargemodelsstrugglewithourEncyclopedic-VQAdataset,ourexperimentsshowthepromiseofaugmentingthemwith
aretrievalmechanism.
PaLM and GPT-3. Models with a larger language un- trieval component shows great potential for predicting de-
derstanding component [11, 10, 17, 40] might be bet- tailedproperties,and(3)sincePaLMandGPT-3workmuch
ter at memorizing detailed properties and extracting infor- better than PaLI, having a strong language understanding
mation from textual knowledge base entries. Therefore, modelisimportantforextractingtherightinformationfrom
we experiment with PaLM 2 [17] (text-bison@001 thefree-formtextKBarticle.
model, available through the PaLM API) and GPT-3 [10] OracleretrievalofKBsection. Finally,tounderstandthe
(text-davinci-003, accessible through the OpenAI importanceofretrievinginformationthatismoreprecisely
API).Botharetrainedonmassiveamountsoftextincluding localizedthananentireWikipediapage, weprovidetothe
Wikipedia. GPT-3 is especially large, with 175B parame- modelstheground-truthsectionwhichsupportstheanswer
ters. Here we feed each model only the text questions Q (Oracle-KBSectioninTab.5). Again,accuracyimproves
as thery cannot consume images. This measures how well to48.8%forPaLI,87.0%forPaLMand82.1%forGPT-3.
our dataset can be solved by language alone. These mod- Thisdemonstratestheimportanceofprovidingexactinfor-
elsreachamodestaccuracy(19.7%,15.5%,Tab.5). While mationtothelanguagecomponent. Asabonus,thesmaller
they only takes textual inputs, they can be made multi- the retrieved document is upon which the model bases its
modal by adding a visual retrieval mechanism, as we do answer,themoreverifiableandinterpretablethisansweris.
inthefollowingsections.
5.3.LargeModelswithVisualRetrieval
5.2.LargeModelswithOracleRetrieval
Lens-basedretrievalofKBarticle. Togobeyondtheora-
Oracle retrieval of subject C. This experiment tests cledemonstrationabove,asproof-of-conceptweproposeto
whether the most difficult aspect of Encyclopedic-VQA is augmentlargemodelswitharealretrievalsystembasedon
recognizingfine-grainedcategoriesandinstancesintheim- GoogleLens[1]. GoogleLensisanimageretrievalsystem
age. Therefore we provide along with the test question its whichindexesahugeamountofwebimages.Givenaquery
corresponding subject category C as an additional textual image,itfindsotherimagesbasedontheirvisualsimilarity
inputtoourmodels(i.e.aspartoftheprompt;Oracle-Sub- andrelevancetoobjectsitrecognizesinthequeryimage. It
ject C in Tab. 5). While PaLI shows small improvements returnsthemostsimilarindexedimagesalongwithanentity
over its non-retrieval version, PaLM and GPT-3 improve prediction based on these top-ranked images. To augment
considerably to 31.0% and 26.9%, respectively. However, oursystem,wesendGoogleLensthequeryimageI toob-
thisisstillratherlowperformance. Weconcludethatdeter- tain its entity prediction. We then find the best matching
miningthefine-grainedcategoryorinstanceisonlypartof KB article for this entity in our knowledge base. Finally,
whatmakesourdatasetdifficult. wefeedtheretrievedKBarticleasprompttoourmodelsas
OracleretrievalofKB	article.	Wenowgoa		stepfurther		inSec.5.2.
-------------------	--------	--------	---	-----------	---	----------	---	---	---
andprovidethefullground-truthWikipediaarticleaboutC Results are shown as Lens - KB Article in Tab. 5. All
tothemodelsintheprompt(Oracle,KBArticleinTab.5). models greatly outperform their non-retrieval augmented
This time, results increase dramatically: 29.7% accuracy versions. PaLMandGPT-3inparticularmorethandouble,
for PaLI, 78.4% for PaLM and 77.4% for GPT-3. This to48.0%and44.9%respectively.
demonstrates that (1) memorizing detailed properties is a Lens-basedretrievalofKBsection. GivenaretrievedKB
hard challenge for vanilla large VLMs, (2) adding a re- article, wenowaimtoselectthemostrelevantKBsection

			Accuracy				5.5.ComparisontoPromptCap
			w/correct	w/incorrect		KBretrieval
							We compare		our retrieval-based			methods	to	Prompt-
LensRetrieval	Overall		retrieval	retrieval		accuracy
							Cap [25],	which	has	strong	performance		on OK-VQA
KBArticle	48.0%		77.7%	21.2%		47.4%
(Tab.4). PromptCapisasystemwithmultiplecomponents.
KBSection	48.8%		82.3%	20.7%		45.6%
---------	-----	---	-----	-----	---	-----	--------	----------	-----	-----	--------	---	------	-----
							It has a	captioning	model	which	consumes		both I	and Q
andgeneratesanimagecaptiontailoredtoanswertheques-
Table6:Attributionforretrieval-augmentedVLMs.Wereport
							tion Q.	This caption	is	then	passed	to GPT-3	as	context
---	---	---	---	---	---	---	-------	------------	---	----	------	--------	---	-------
modelaccuracyforPaLM[17]conditionalonretrievalresults.Re-
trievalsuccessratesaresimilarforarticlesandsections,butPaLM toanswerQ. Additionally,PromptCapperformsin-context
hasbetteraccuracywhenaugmentedwithsections. learning: for a given test question, it uses the CLIP em-
							beddings	[39]	of Q and	I to	find the	32 nearest		training
---	---	---	---	---	---	---	---------	-----------------------------------------	--------	----	--------	----------	---	--------
							examples.	Thenitincludestheir32correspondingPrompt-
Capcaptions,questionsandanswersintheGPT-3prompt.
withinitbasedonQ.		TodosowequeryPaLMwithaspe-
-----------------	---	--------------------------	---	---	---	---	---	---	---	---	---	---	---	---
cialpromptfeedingoneKBsectionatatime,alongwithQ In[25]theunderlyingLLMwasGPT-3. Inthisexperiment
while asking ‘can the answer to this question be found in wealsoapplyPromptCapwithPaLMandPaLI.Resultsin
thistext?’. WeretainallsectionsforwhichPaLManswers Tab. 5 show that PromptCap indeed does better than using
‘Yes’ (usually only one), and consider their concatenation PaLI, PaLM, or GPT-3 alone. However, it still performs
							substantially	worse	than	our retrieval-augmented			methods.
-------------------------------	---	---	---	---------------------	---	---	-------------	-----	----	-----------------------	---	---	--------	---
asthefinalretrieved‘KBsection’.				Finally,wefeedthisre-
trievedsectionintheprompttoallmodels(PaLI,PaLMor Thisfurtherconfirmsthehardchallengeourdatasetposes.
GPT-3)asinSec.5.2.
5.6.Multi-answerandtwo-hopquestions
Results	are shown	as	Lens -	KB Section	in	Tab. 5. Per-
-------	---------	---	------	----------	---	------------	---	---	---	---	---	---	---	---
formance is now even higher for PaLI (28.1%) and PaLM To evaluate multi-answer questions, we convert the
							model prediction		into	a set	of strings	and	compute	the
--------	-------	-----	----	--------	----------	--------	---------------------------	---	----	-----	----------	----	-------	-------
(48.8%),	whereas	GPT-3	does	not seem	to benefit	further.
							intersection-over-the-union			(IoU)	between	this	set	and the
Thesenumbersareroughlyhalfwaybetweenversionswith-
outretrievalandwithoracleretrieval. Weconcludethatre- set of ground-truth answers. If IoU >= 0.5 then we con-
							sider that	prediction	as	correct.	If not,	then	we use	BEM
---	---	---	---	---	---	---	----------	----------	---	--------	-------	----	------	---
trievalaugmentationworksinpractice,andthatourdataset
leavessignificantheadroomforfutureresearchonretrieval- to determine the equivalence of the prediction list string
							and the	ground-truth	list.	For	two-hop	questions,		we use
---	---	---	---	---	---	---	-------	------------	-----	---	-------	----------	---	------
augmentedVLMs.
							BEM as	described	for	single-hop	templated		and automati-
----------	----------	---	---	--------	------	--------	------	---------	---	----------	---------	---	-------------	---
CLIP-based	retrieval.		KAT	[22] and	REVIVE	[34] are
callygeneratedquestions.
two retrieval-augmented			VQA	systems	which	use frozen
-----------------------	----------	---	-------	----------	-----	-----------	----	---------	---	-------	---------	-------	---	-----
							PaLI	with Lens	KB	Section	Retrieval	obtains	an	accu-
CLIP [39]	embeddings	to	perform	retrieval.		In Appendix
racyof9.2%formulti-answerquestionsand14.7%fortwo-
C we explore	there	whether	KAT	and	REVIVE	could po-
------------	-----	-------	---	---	------	---------	--------------	---	----------	----	----	----------	---	---------
							hop questions.		Similarly,	PaLM	with	KB Section		retrieval
tentiallysucceedonEncyclopedic-VQA.
							achieves	33.6%	and 22.8%	for	multi-answer		and	two-hop
---	---	---	---	---	---	---	----------	-------------	---------	-----	------------	-----	---	-------
							questions,	respectively.		GPT-3	obtains	32.1%	for	multi-
answerand18.7%fortwo-hopquestions,inbetweenPaLI
5.4.AttributionforRetrieval-Augmentedmodels
andPaLM.Thesemodestperformances,especiallyontwo-
hopquestions,confirmthechallengeofourproposedtasks,
The attribution		annotations		of our dataset		enable mea-
---------------	---	-----------	---	--------------	---	-----------	---	---	---	---	---	---	---	---
andhighlighttheusefulnessofEncyclopedic-VQAtomea-
suringwhetherretrieval-augmentedmodelsgivethecorrect
							sure progress	on	designing	retrieval		mechanisms	for	such
------------------------	---	---	---------------------------	---	---	---	-------------	---	---------	---------	---	----------	---	----
answerfortherightreason.			Tab.6measuresattributionfor
complextypesofquestions.
thePaLMmodelswithLensretrieval.
First,	we observe	that	Google	Lens	retrieves	the cor-
------	----------	----	------	----	---------	--------	---	---	---	---	---	---	---	---
6.Conclusions
rect KB	Article 47.4%		of the	time. Furthermore,		we re-
-------	-------------	---	------	------------------	---	------	---	---	---	---	---	---	---	---
port 45.6% retrieval accuracy at the finer level of a KB We introduced Encyclopedic-VQA, a large-scale VQA
Section. Thisdemonstratesthatgivenacorrectlyretrieved dataset about detailed properties of fine-grained categories
KBArticle,oursystemalmostalwaysfindsthecorrectSec- and instances, which includes a knowledge base with an-
tion within it. More importantly, if the retrieved KB Sec- swerattributions. Wedemonstratedthatourdatasetistruly
tionisincorrect,accuracydropsdrasticallyto20.7%,close difficultforstandardVLMs. Additionally,weshowedwith
tothenon-augmentedvariant(bothwhenusingPaLM).In both an oracle experiment and a prototype automatic sys-
contrast, if the correct KB Section is found, accuracy is tem that augmenting these models with a retrieval compo-
82.3%.Thus,retrievingthecorrectdocument,andevenbet- nent substantially improves results, yet leaving headroom
terthecorrectsection,isessentialtoperformingwellonour for even further improvements. Therefore our dataset en-
dataset.							ablesfutureresearchonretrieval-augmentedVLMs.
--------	---	---	---	---	---	---	---------------------------------------------	---	---	---	---	---	---	---

References [12] Jannis Bulian, Christian Buck, Wojciech Gajewski, Ben-

							jamin	Boerschinger,		and Tal	Schuster.	Tomayto,		tomahto.
[1] GoogleLens. https://lens.google.com-Webinterfaceavail- beyondtoken-levelanswerequivalenceforquestionanswer-
ableathttps://images.google.com.				8,13
--------------------------------	---------	-----	---------	---------	------------	---	--------------	---	------------------------------	---	---	---	---	--------
							ingevaluation.		arXivpreprintarXiv:2202.07654,					2022. 7,
[2] Google	universal	image	embedding	challenge	leaderboard.
13
https://www.kaggle.com/competitions/google-universal- [13] Soravit Changpinyo, Doron Kukliansky, Idan Szpektor, Xi
image-embedding/leaderboard. 14 Chen, NanDing, andRaduSoricut. Allyoumayneedfor
[3] Huggingface. https://huggingface.co/sentence- VQAareimagecaptions. InNAACL,2022. 4,5
transformers/multi-qa-MiniLM-L6-cos-v1. 7 [14] XiChen,XiaoWang,SoravitChangpinyo,AJPiergiovanni,
[4] inat challenge 2021 - fgvc8 - leaderboard. Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam
https://www.kaggle.com/competitions/inaturalist- Grycner,BasilMustafa,LucasBeyer,AlexanderKolesnikov,
							Joan	Puigcerver,	Nan	Ding,	Keran	Rong,	Hassan	Akbari,
-----------------	---	---	---	---	---	---	----	-----------	---	-----	-----	-----	------	-------
2021/leaderboard.		14
[5] WikimediaCommons. https://commons.wikimedia.org. 4 GauravMishra,LintingXue,AshishThapliyal,JamesBrad-
							bury,	Weicheng	Kuo,	Mojtaba	Seyedhosseini,			Chao Jia,
---	---	---	---	---	---	---	-----	--------	----	-------	--------------	---	---	---------
[6] ArmenAghajanyan,BernieHuang,CandaceRoss,Vladimir
							Burcu	Karagol	Ayan,	Carlos	Riquelme,		Andreas	Steiner,
---	---	---	---	---	---	---	-----	-------	-----	------	---------	---	-------	--------
Karpukhin,HuXu,NamanGoyal,DmytroOkhonko,Man-
							Anelia	Angelova,	Xiaohua		Zhai, Neil	Houlsby,		and Radu
---	---	---	---	---	---	---	------	---------	-------	---	----------	--------	---	--------
darJoshi,GargiGhosh,MikeLewis,andLukeZettlemoyer.
							Soricut.	PaLI:	A Jointly-Scaled			Multilingual	Language-
---	---	---	---	---	---	---	--------	-----	----------------	---	---	------------	---------	---
CM3:Acausalmaskedmultimodalmodeloftheinternet.In
							ImageModel.		arXiv,2022.	1,2,3,7,8
-----------------	------------	----------	-------------	-------------	-----------	---------	-----------------------------	---------	------------	----------	--------------	---------	--------	---------
arXiv,2022.		3
							[15] Yang	Chen,	Hexiang	Hu, Yi	Luan,	Haitian	Sun,	Soravit
[7] Jean-Baptiste		Alayrac,	Jeff Donahue,		Pauline	Luc, An-
							Changpinyo,		Alan Ritter,		and Ming-Wei		Chang.	Can
toine	Miech,	Iain Barr,	Yana	Hasson, Karel	Lenc,	Arthur
							pre-trained		vision and	language		models	answer	visual
Mensch,	Katie	Millican,	Malcolm	Reynolds,	Roman	Ring,
							information-seekingquestions?				InarXiv,2023.		4
Eliza	Rutherford,	Serkan	Cabi,	Tengda Han,	Zhitao	Gong,
							[16] Yen-Chun	Chen,	Linjie	Li,	Licheng	Yu, Ahmed		El Kholy,
Sina	Samangooei,	Marianne	Monteiro,	Jacob	Menick,	Se-
							Faisal	Ahmed,	Zhe	Gan,	Yu Cheng,	and	Jingjing	Liu.
bastian	Borgeaud,	Andrew	Brock,	Aida	Nematzadeh,	Sa-
							UNITER:	UNiversal		Image-TExt	Representation			Learning.
hand	Sharifzadeh,	Mikolaj	Binkowski,		Ricardo	Barreira,
							InECCV,2020.		4
Oriol	Vinyals,	Andrew	Zisserman,	and	Karen	Simonyan.
							[17] Aakanksha		Chowdhery,	Sharan		Narang,	Jacob	Devlin,
Flamingo:aVisualLanguageModelforFew-ShotLearning.
							Maarten	Bosma,	Gaurav		Mishra,	Adam	Roberts,	Paul
-------------	------	---------	--------	------	---	--------	-------	---------	------	------	-------	------	--------	------
NeurIPS,2022.		1,2,3
							Barham,	Hyung	Won	Chung,	Charles		Sutton,	Sebas-
[8] Stanislaw	Antol,	Aishwarya	Agrawal,	Jiasen	Lu,	Margaret
							tian	Gehrmann,	Parker		Schuh,	Kensen	Shi,	Sasha
Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker
Parikh. VQA: visual question answering. In ICCV, 2015. Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran,
3,13
							EmilyReif,		NanDu,	BenHutchinson,		ReinerPope,		James
-------------	---------	------	---	-------	------	---------	----------	-----	-------	--------------	-------	-----------	---	--------
[9] Sebastian	Borgeaud,	Arthur		Mensch,	Jordan	Hoffmann,
							Bradbury,	Jacob	Austin,		Michael	Isard,	Guy	Gur-Ari,
TrevorCai,ElizaRutherford,KatieMillican,Georgevanden Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay
Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier
Clark,DiegodeLasCasas,AureliaGuy,JacobMenick,Ro- Garcia,VedantMisra,KevinRobinson,LiamFedus,Denny
manRing,TomHennigan,SaffronHuang,LorenMaggiore,
							Zhou,	Daphne	Ippolito,	David	Luan,		Hyeontaek	Lim,
---	---	---	---	---	---	---	-----	------	---------	-----	-----	---	---------	----
ChrisJones,AlbinCassirer,AndyBrock,MichelaPaganini,
							Barret	Zoph,	Alexander	Spiridonov,		Ryan	Sepassi,	David
---	---	---	---	---	---	---	------	-----	---------	-----------	---	----	--------	-----
GeoffreyIrving, OriolVinyals, SimonOsindero, KarenSi- Dohan, Shivani Agrawal, Mark Omernick, Andrew M.
monyan,JackW.Rae,ErichElsen,andLaurentSifre. Im- Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat,
provinglanguagemodelsbyretrievingfromtrillionsofto- AitorLewkowycz, EricaMoreira, RewonChild, Oleksandr
kens.	arXiv:2112.04426,2022.			1,4
-----	----------------------	---	---	---	---	---	--------	---------	---	------------	---	-----	------	-----
							Polozov,	Katherine		Lee, Zongwei		Zhou,	Xuezhi	Wang,
[10] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Sub- Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta,
biah,JaredDKaplan,PrafullaDhariwal,ArvindNeelakan- JasonWei,KathyMeier-Hellstern,DouglasEck,JeffDean,
tan,PranavShyam,GirishSastry,AmandaAskell,etal.Lan- Slav Petrov, and Noah Fiedel. PaLM: Scaling Language
guagemodelsarefew-shotlearners. Advancesinneuralin- ModelingwithPathways. arXiv,2022. 5,7,8,9,14
formationprocessingsystems,33:1877–1901,2020.
						7,8	[18] Hyung	Won	Chung,	Le Hou,	Shayne		Longpre,	Barret
---	---	---	---	---	---	---	----------	---	------	-------	------	---	--------	------
[11] TomB.Brown,BenjaminMann,NickRyder,MelanieSub- Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang,
biah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakan- Mostafa Dehghani, Siddhartha Brahma, et al. Scaling
tan, Pranav Shyam, Girish Sastry, Amanda Askell, Sand- instruction-finetuned language models. arXiv preprint
							arXiv:2210.11416,2022.			5,14
-------------	---	-------------------	---	--------	--------	---	----------------------	---	---	----	---	---	---	---
hini Agarwal,		Ariel Herbert-Voss,		Gretchen	Krueger,	Tom
Henighan,RewonChild,AdityaRamesh,DanielM.Ziegler, [19] Matthew Dunn, Levent Sagun, Mike Higgins, V. Ugur
JeffreyWu,ClemensWinter,ChristopherHesse,MarkChen, Guney, Volkan Cirik, and Kyunghyun Cho. SearchQA: A
Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, newQ&Adatasetaugmentedwithcontextfromasearchen-
JackClark,ChristopherBerner,SamMcCandlish,AlecRad- gine. InarXiv,2017. 4
ford,IlyaSutskever,andDarioAmodei. LanguageModels [20] HaoyuanGao, JunhuaMao, JieZhou, ZhihengHuang, Lei
areFew-ShotLearners. arXiv,2020. 8 Wang,andWeiXu. Areyoutalkingtoamachine? Dataset

andmethodsformultilingualimagequestion. InNeurIPS, [34] YuanzeLin,YujiaXie,DongdongChen,YichongXu,Chen-

2015.	3						guangZhu,andLuYuan.			Revive:Regionalvisualrepresen-
							tation matters	in	knowledge-based		visual question	answer-
[21] YashGoyal,TejasKhot,DouglasSummers-Stay,DhruvBa-
tra, and Devi Parikh. Making the V in VQA matter: Ele- ing. arXivpreprintarXiv:2206.01201,2022. 9,13,14
vating the role of image understanding in Visual Question [35] Mateusz Malinowski and Mario Fritz. A multi-world ap-
Answering. InCVPR,2017. 3,7,13 proachtoquestionansweringaboutreal-worldscenesbased
InNeurIPS,2014.
[22] Liangke Gui, Borui Wang, Qiuyuan Huang, Alex Haupt- onuncertaininput. 3
mann, Yonatan Bisk, and Jianfeng Gao. KAT: A Knowl- [36] KennethMarino,XinleiChen,DeviParikh,AbhinavGupta,
edgeAugmentedTransformerforVision-and-Language. In and Marcus Rohrbach. KRISP: Integrating Implicit and
NAACL,2022. 1,4,7,9,13,14 Symbolic Knowledge for Open-Domain Knowledge-Based
[23] KelvinGuu,KentonLee,ZoraTung,PanupongPasupat,and VQA. InProc.CVPR,2021. 1,4
Ming-WeiChang.REALM:Retrieval-AugmentedLanguage [37] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and
ModelPre-Training. InProc.ICML,2020. 1,4 RoozbehMottaghi. OK-VQA:Avisualquestionanswering
[24] Yaru	Hao,	Haoyu Song,	Li	Dong, Shaohan	Huang,	Zewen
------------------------------------	----	-----------	---	-------------	------	--------	----------	---------	--------	----------	---	-----------
							benchmark	requiring	external	knowledge.	In	CVPR, 2019.
Chi,WenhuiWang,ShumingMa,andFuruWei.						Language	1,2,3,7,13
modelsaregeneral-purposeinterfaces. InarXiv,2022. 1,2, [38] TodorMihaylov,PeterClark,TusharKhot,andAshishSab-
3
							harwal.	Can a	suit of armor	conduct	electricity?	a new
---	---	---	---	---	---	---	-------	-----	-------------	-------	------------	-----
[25] YushiHu,HangHua,ZhengyuanYang,WeijiaShi,NoahA
datasetforopenbookquestionanswering.InEMNLP,2018.
Smith,	and	Jiebo Luo.	Promptcap:	Prompt-guided		task-	4
---------------------	---	----------	------------------------------	-------------	-----	----------	------------------	-------	-------------	--------	--------------	--------------
awareimagecaptioning.			arXivpreprintarXiv:2211.09699,
							[39] Alec Radford,		Jong Wook	Kim,	Chris Hallacy,	Aditya
2022.	7,9
							Ramesh,	Gabriel	Goh, Sandhini		Agarwal,	Girish Sastry,
[26] Ziniu	Hu,	Ahmet	Iscen,	Chen Sun,	Zirui	Wang, Kai-
							Amanda	Askell,	Pamela	Mishkin,	Jack Clark,	Gretchen
Wei Chang, Yizhou Sun, Cordelia Schmid, David A. Krueger, and Ilya Sutskever. Learning transferable visual
Ross, and Alireza Fathi. REVEAL: Retrieval-Augmented modelsfromnaturallanguagesupervision. InICML,2021.
Visual-Language		Pre-Training		with Multi-Source		Multi-
---------------	---	------------	---	-----------------	---	------	---	---	---	---	---	---
3,9,13
modalKnowledgeMemory.				InCVPR,2023.	1,4,7
---------------------	---	---	---	------------	-----	---	---------------	------------------	---	---	----------	-----------
							[40] JackW.Rae,	SebastianBorgeaud,			TrevorCai,	KatieMilli-
[27] Aman Jain, Mayank Kothyari, Vishwajeet Kumar, Preethi can,JordanHoffmann,FrancisSong,JohnAslanides,Sarah
Jyothi,GaneshRamakrishnan,andSoumenChakrabarti.Se- Henderson, Roman Ring, Susannah Young, Eliza Ruther-
lect,	substitute,	search:	A	new benchmark	for	knowledge-
-----	-----------	-------	---	-------------	---	----------	---	---	---	---	---	---
ford,TomHennigan,JacobMenick,AlbinCassirer,Richard
augmentedvisualquestionanswering.				InSIGIR,2021.		2,3,
---------------------------------	---	---	---	-------------	---	----	--------------	---	------------------	---	---------	----------
							Powell, George		van den Driessche,		Lisa Anne	Hendricks,
4,5,7
							Maribeth	Rauh,	Po-Sen Huang,		Amelia Glaese,	Johannes
---	---	---	---	---	---	---	--------	-----	-------------	---	--------------	--------
[28] ChaoJia,YinfeiYang,YeXia,Yi-TingChen,ZaranaParekh, Welbl, Sumanth Dathathri, Saffron Huang, Jonathan Ue-
HieuPham,QuocV.Le,Yun-HsuanSung,ZhenLi,andTom sato, John Mellor, Irina Higgins, Antonia Creswell, Nat
Duerig.Scalingupvisualandvision-languagerepresentation
							McAleese,	Amy	Wu, Erich	Elsen,	Siddhant	Jayakumar,
---------------------------------	---	---	---	------------	---	---	---------	---	---------	------	--------	----------
learningwithnoisytextsupervision.				InICML,2021.		3
ElenaBuchatskaya,DavidBudden,EsmeSutherland,Karen
[29] MandarJoshi,EunsolChoi,DanielS.Weld,andLukeZettle- Simonyan, MichelaPaganini, LaurentSifre, LenaMartens,
moyer. TriviaQA: A large scale distantly supervised chal- Xiang Lorraine Li, Adhiguna Kuncoro, Aida Nematzadeh,
lengedatasetforreadingcomprehension. InACL,2017. 4 Elena Gribovskaya, Domenic Donato, Angeliki Lazaridou,
[30] UrvashiKhandelwal,OmerLevy,DanJurafsky,LukeZettle- ArthurMensch,Jean-BaptisteLespiau,MariaTsimpoukelli,
moyer,andMikeLewis. Generalizationthroughmemoriza- NikolaiGrigorev,DougFritz,ThibaultSottiaux,MantasPa-
tion: Nearestneighborlanguagemodels. InICLR,2020. 1, jarskas, Toby Pohlen, Zhitao Gong, Daniel Toyama, Cy-
4							priendeMassond’Autume,YujiaLi,TayfunTerzi,Vladimir
---	---	---	---	---	---	---	--------------------------------------------------	---	---	---	---	---
Mikulik,IgorBabuschkin,AidanClark,DiegodeLasCasas,
[31] Tushar	Khot,	Peter	Clark,	Michal Guerquin,	Peter	Jansen,
-----------	-----	-----	------	----------------	-----	-------	---	---	---	---	---	---
and Ashish Sabharwal. QASC: A dataset for question an- Aurelia Guy, Chris Jones, James Bradbury, Matthew John-
sweringviasentencecomposition. InAAAI,2020. 4 son, Blake Hechtman, Laura Weidinger, Iason Gabriel,
[32] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Ui- WilliamIsaac,EdLockhart,SimonOsindero,LauraRimell,
							Chris Dyer,	Oriol	Vinyals,	Kareem	Ayoub,	Jeff Stanway,
---	---	---	---	---	---	---	-----------	-----	--------	------	------	-------------
jlings,IvanKrasin,JordiPont-Tuset,ShahabKamali,Stefan
							Lorrayne	Bennett,	Demis	Hassabis,	Koray	Kavukcuoglu,
-----------------------	------	--------	---	---------------------------	--------	--------	-------------------	--------	-------	---------	-------	------------
Popov,	Matteo	Malloci,	Tom	Duerig, and	Vittorio	Ferrari.
							and GeoffreyIrving.		Scaling	Language	Models:	Methods,
TheOpenImagesDatasetV4:				Unifiedimageclassification,
object detection, and visual relationship detection at scale. Analysis&InsightsfromTrainingGopher. arXiv,2022. 8
IJCV,2020. 4 [41] ColinRaffel,NoamShazeer,AdamRoberts,KatherineLee,
SharanNarang,MichaelMatena,YanqiZhou,WeiLi,,and
[33] Patrick	Lewis,	Ethan	Perez,	Aleksandra	Piktus,	Fabio
------------	------	-----	------	----------	-------	-----	---	---	---	---	---	---
Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich PeterJLiu. Exploringthelimitsoftransferlearningwitha
Ku¨ttler, Mike Lewis, Wen tau Yih, Tim Rockta¨schel, Se- unifiedtext-to-texttransformer. JMLR,2021. 13
bastian Riedel, and Douwe Kiela. Retrieval-Augmented [42] Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and
Generation for Knowledge-Intensive NLP Tasks. In Proc. PercyLiang.SQuAD:100,000+questionsformachinecom-
NeurIPS,2020.		1,4					prehensionoftext.		InEMNLP,2016.		4
-------------	---	---	---	---	---	---	-----------------	---	-------------	---	---	---

[43] NilsReimersandIrynaGurevych. Sentence-bert: Sentence [60] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, embeddingsusingsiamesebert-networks. InEMNLP,2019. Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang,

7								BoxinLi,ChunyuanLi,CeLiu,MengchenLiu,ZichengLiu,
[44] MengyeRen,RyanKiros,andRichardS.Zemel. Exploring YumaoLu,YuShi,LijuanWang,JianfengWang,BinXiao,
modelsanddataforimagequestionanswering. InNeurIPS, ZhenXiao,JianweiYang,MichaelZeng,LuoweiZhou,and
2015.	4							PengchuanZhang.	Florence: ANewFoundationModelfor
------------	------	----------	---	------------------	---	---	----------	---------------	--------------------------------
								ComputerVision.	arXiv,2021. 3
[45] Eleanor	Rosch.	Principles		of categorization.			In Eleanor
RoschandB.B.Lloyd,editors,CognitionandCategoriza-
tion,pages27–48.LawrenceErlbaumAssociates,1978.							2
-----------------------------------------------	--------	------	-------	-----------	---	-----------	------	---	---
[46] Dustin	Schwenk,	Apoorv		Khandelwal,		Christopher	Clark,
Kenneth	Marino,	and	Roozbeh	Mottaghi.		A-OKVQA:	A
benchmarkforvisualquestionansweringusingworldknowl-
edge.	InECCV,2022.		2,3,7,13
------------------------	------------	---------------------------------------	------------	------------	--------------------	-------	---------	---	---
[47] Sanket	Shah,	Anand	Mishra,		Naganand	Yadati,	and
Partha	Pratim	Talukdar.	KVQA:		Knowledge-aware		visual
questionanswering.			InAAAI,2019.		2,3,7
[48] Krishna	Srinivasan,		Karthik	Raman,	Jiecao	Chen,	Michael
Bendersky,andMarcNajork.				Wit:	Wikipedia-basedimage
text	dataset	for multimodal		multilingual		machine	learning.
SIGIR,2021.		3,4
[49] AndyStirling.		Ageneralframeworkforanalysisngdiversity
in science,		technology,	and	society.	Journal	of	the Royal
SocietyInterface,2007.			7
[50] AlonTalmorandJonathanBerant.Thewebasaknowledge-
baseforansweringcomplexquestions.					InNAACL,2018.		4
---------------------------------	---	---	---	---	-------------	---	---	---	---
[51] GrantvanHorn,ElijahCole,SaraBeery,KimberlyWilber,
SergeBelongie,andOisinMacAodha.					BenchmarkingRep-
-------------------------------	---	--------	-----------	-----	----------------	------------	---	---	---
resentation		Learning	for Natural	World	Image	Collections.
InCVPR,2021.		2,4
[52] PeterVickers,NikolaosAletras,EmilioMonti,andLoicBar-
rault.	InFactuality:		EfficientIntegrationofRelevantFacts
---------------------------	-------------	---	-----------------------------------	-----------	---	---	---	---	---
forVisualQuestionAnswering.				InACL,2021.		1,4
[53] PengWang,QiWu,ChunhuaShen,AnthonyDick,andAn-
tonvandenHengel.			FVQA:Fact-basedvisualquestionan-
----------------	----------------------	------	--------------------------------	----	-----------	---	-------	---	---
swering.	IEEETrans.onPAMI,2018.				3,7
[54] Jason	Wei,	Xuezhi	Wang,	Dale	Schuurmans,		Maarten
Bosma,BrianIchter,FeiXia,EdChi,QuocLe,andDenny
Zhou.	Chain-of-Thought			Prompting	Elicits	Reasoning	in
--------------------	----------------	------	---------------	---------	-------	---------	-------	---	---
LargeLanguageModels.			InNeurIPS,2022.			14
[55] Johannes	Welbl,	Pontus	Stenetorp,		and	Sebastian	Riedel.
Constructingdatasetsformulti-hopreadingcomprehension
acrossdocuments.			InACL,2018.		4
----------------	---	----------	-----------	--------	-------	------	-----	---	---
[56] T. Weyand,		A. Araujo,	B.	Cao, and	J. Sim.	Google	land-
marksdatasetv2-alarge-scalebenchmarkforinstance-level
recognitionandretrieval.			InCVPR,2020.			2,4
------------------------	---	---	------------	---	---	---	---	---	---
[57] ZhengyuanYang,ZheGan,JianfengWang,XiaoweiHu,Yu-
maoLu,ZichengLiu,andLijuanWang.						Anempiricalstudy
-------------------------------------	---	---	---	---	---	----------------	---	---	---
ofgpt-3forfew-shotknowledge-basedvqa.						InAAAI,2022.
1,2,3
[58] Zhilin	Yang,	Peng	Qi, Saizheng		Zhang,	Yoshua	Bengio,
-------------------------------	------	------------	--------------	-------------	------------	-----------	-----------	---	---
William	Cohen,	Ruslan	Salakhutdinov,		and	Christopher	D.
Manning.		HotpotQA:	A	dataset	for diverse,		explainable
multi-hopquestionanswering.				InEMNLP,2018.			3,4,6
[59] Licheng	Yu,	Eunbyung	Park,	Alexander		C.	Berg, and
Tamara	L.	Berg. Visual	madlibs:		Fill in	the blank	image
generationandquestionanswering.					InICCV,2015.		3

A.Correctnessofourdataset astrictevaluationmeasureandbecausetheirquestionsare

							more	open-ended		(they	are not supported		by a controlled
We rely	on	iNat21	and GLDv2	to have	clearly	identifi-
							knowledgebase,			Sec.4),	thosedatasets	collect10		ground-
ablecategoriesonthetestimagesofourdataset,andtoob-
truthanswersperquestion,tocoversomeofthevariability
taintheirground-truthlabels.Furthermore,weonlyusecat- in answer formulation. This then enables a more relaxed
egorieswhichunambiguouslymaptoWikipediaarticlesby
							evaluation,		as the	model	answer need	only	match	some of
------------------------------------------------	---	---	---	---	---	---	-----------	---------------	------	-----	--------------	----	-------------	-------
explicitlyseekingforone-to-onemappings(Sec.4.1).						The
							the	10 ground-truth		ones	(three matches		for a perfect	an-
mainremainingpossiblesourceoferroriswhetheraques-
swer). Nevertheless,itisinstructivetoverifywhetherexact
tion can	be answered		given	the corresponding		Wikipedia
--------	-----------	---	-----	-----------------	---	---------	---	---	---	---	---	---	---	---
matchingwouldworkonourdataset,wherewehaveasin-
article,	and whether		the recorded	ground-truth		answer is
--------	-----------	-----	------------	--------------	---	---------	---	------------	---	-------	--------	--------	--------	------
							gle	ground-truth		answer.	We start	from the	publicly	avail-
indeed	correct.	Human	annotators	systematically		vali-
ableexactmatchingimplementationof[22]andincludead-
date these	aspects	for	every	single question+answer		pair
----------------	-------	---------------------------------	-----	----------------------	---	----	--------------------------------------------	---	---------	-----	--------	------	-----------	---
							ditionalrelaxationsfornumbercomparisons(e.g.						numbers
(Sec.4.2and4.3).		Yet,weperformhereanadditionaluser
							over	100	- usually	years	- may be	off by	one, ranges	are
studytogetannumericalestimateofqualityinthissense.
							correct	if	they	partially	overlap, etc).	We	find that	exact
---	---	---	---	---	---	---	-------	---	----	---------	--------------	---	---------	-----
In this study, we randomly sample 100 questions from matchjudgementsareequivalenttohumanjudgementsonly
thetestset.Thenweasksixexpertstoeachanswer50ques-
							in	68% of	the cases.	Where	they	differ, exact	match	is al-
---	---	---	---	---	---	---	---	------	----------	-----	----	-------------	-----	------
tionseach,giventhecorrespondingWikipediapageasref-
							ways	overly	strict,	rejecting	answers	that	a human	would
---	---	---	---	---	---	---	----	------	-------	---------	-------	----	-------	-----
erence. Theseexpertwerenotinvolvedintheoriginaldata judge as correct. This demonstrates that exact matching
collectionprocess.Thisprocessresultsinthreeanswersper
doesnotworkwellforourdataset,andjustifiesourchoice
question. If the majority of the 3 expert answers matches of BEM as the evaluation measure to check whether pre-
ourcollectedground-truthaccordingtoBEM[12],wecon-
							dictedmodelanswersmatchground-truthonesinSec.							5.
----------	--------	---	-------------	-----	---	---------	---------------------------------------------	---	---	---	---	---	---	---
sider that	question	to	be answerable	given	the	Wikipedia
page, andourground-truthto becorrect. Wefind thatthis C. CLIP retrieval in existing retrieval-
holdsformostofourquestions(86%).
augmentedVQAsystems[22,34]
To put	this number		in context,	we also	estimate	the an-
------	-----------	---	-----------	-------	--------	-------	---	---	---	---	---	---	---	---
swerabilityofA-OKVQA[46]withrespecttotheirevalua- KAT[22]andREVIVE[34]aretworetrieval-augmented
tionmetric.A-OKVQAfollowspreviousVQAdatasetsand VQA systems which use frozen CLIP [39] embeddings to
provides10ground-truthanswersperquestion. Apredicted performretrieval. Morespecifically,theyfirstencodetheir
answer is counted as correct if it matches with 3 ground- text-only knowledge base (extracted from Wikidata) using
truth answers (exact string matching [8, 46], after normal- thelanguagetowerofCLIP.Attesttime,givenaVQAques-
izationsuchasremovingpunctuationandarticles,convert- tion,theyencodeitsimageI withCLIPandthencompare
ingnumberstodigits,etc.). Sowecanconsideraquestion itsembeddingtotheknowledgebaseembeddingstoretrieve
is answerable if there are at least 3 equivalent answers out the most similar entries. The top few most similar entries
ofthe10.Wefindthat86%oftheA-OKVQAquestionsare are then passed on to a T5 model [41] which is trained to
answerable.
							producethefinalanswer.				Notehowcorrectretrievaliscru-
---	---	---	---	---	---	---	----------------------	---	---	---	-----------------------------	---	---	---
To conclude, this user study demonstrates that our cialtoperformwellonEncyclopedic-VQA,astheretrived
Encyclopedic-VQAdatasetisofveryhighquality. knowledge base entries should contain the answer for the
							overall	system		to succeed.	Hence,	we now	test the	CLIP-
---	---	---	---	---	---	---	-------	------	---	-----------	------	------	--------	-----
B.QualityofBEM[12]evaluationmeasure based retrieval component of KAT [22] and REVIVE [34]
inisolation,toestimatewhethertheywouldbeabletosuc-
We want	to	understand	how	well the	BERT	Matching
-------	---	----------	---	--------	----	--------	---	---	---	---	---	---	---	---
ceedonEncyclopedic-VQA.
(BEM)[12]evaluationmeasuremirrorshumanjudgments. WerepresentWikipediaarticlesastheCLIPembedding
To do so	we build	on	the user	study above	and	ask an ex-
--------	--------	---	--------	-----------	---	----------	---	-----	---------	-----------	-------	-------------	---	-------
							of	their	title and	description	strings	concatenated.		At test
pert human to judge whether each answer from the user time, given a VQA question, we use CLIP to embed the
studymatchesthecollectedground-truthanswerornot.The
judgehasaccesstoboththequestionQandthecorrespond-
ing Wikipedia		page.	We find	that the	BEM	judgements					Recall
-------------	----------	-----	-------	-----------------	---	------------	---	------	---	---	------	---	---	---
equal human	judgements		in	the vast majority		of the cases
								Method		@1	@5	@10	@20
(96%). BEMisastricterjudgein3%ofthecases,thehu-
man in 1%. This demonstrates that BEM correlates very CLIP[39] 3.3% 7.7% 12.1% 16.5%
highlywithhumanjudgementofcorrectnessofananswer, Lens[1] 47.4% 62.5% 64.7% 65.2%
confirmingwhatreportedin[12]forotherdatasets.
Exact Match. Many VQA datasets perform exact string Table7: RecallresultsforCLIPandLens.Wereporttherecall
inretrievingtherightKBarticlewithinthetop-Kdocuments.
matchingfortheirevaluation[8,21,37,46].						Becauseitis
---------------------------------------	---	---	---	---	---	-----------	---	---	---	---	---	---	---	---

imageItoformthequeryforretrieval,proceedingasin[22, E.DatasetStatistics

34]. Results	in	Tab. 7 demonstrate		that recall	is	low: the
								In Table	8 we	provide	a more	detailed	overview	of the
correctWikipediaarticlecorrespondingtothesubjectofthe
							datasetstatistics.			FormoredetailsseeSection4ofthemain
--------	------------	------	--------------	----	----	------	------------------	---	---	----------------------------------	---	---	---	---
question	is retrieved	in the	first position	only	3.3%	of the
paper.
time. Even	within	the top-20	results,	retrieval	accuracy	is
-----------------	------	------------------------------	--------	---------	--------	---	---	---	---	---	---	---	---	---
stillonlyat16.5%.		ThissuggestthatKAT/REVIVEwould
F.Promptsfordatasetcreation
notworkwellonourdataset.
Lens retrieval. In Sec. 5.3 we use Google Lens to iden- AsdiscussedinSection4ofthepaper,weusedaFLAN
tify the subject C of a VQA question in our datatset, i.e. [18] version of the PaLM [17] model to (1) rephrase au-
a iNat21 or GLDv2 category, which works well (Tab. 7). tomatically generated single-hop questions (both single-
Lens is an image-based system which indexes billions of answer and multi-answer), and to (2) combine two single-
images on	the web	and retrieves		relevant	ones	based on
---------	-------	-------------	---	--------	----	--------	---	---	---	---	---	---	---	---
hopquestionsintoatwo-hopquestion.Generally,wefound
their visual similarity. Hence, it has a greater chance to that the model tends to work better when fed with very
findawebimageresemblingaVQAtestimage,thanwhen specific(andsometimesredundant)instructions,combined
restricting the search only to Wikipedia. Attached to the withchain-of-thoughtprompts[54]wherepossible. Addi-
image is often various meta-data which enables to deter- tionally,weaddeddifficultcasesasexamplesintheprompt,
minethenameofthesubjectcategory(whichwethenuseto
							to	guide	the model’s	behavior		in more	challenging	situa-
---------------------------	---	-------------	-----------------------	-------	---------	-----	------	-----	-----------	--------	---	-------	-----------	------
findtherightWikipediapage).			Recognizinglandmarksand				tions.
species from	the	natural world	are	typical	use-cases	so we
F.1.Rephrasingautomaticallygeneratedsingle-hop
can expect	Lens	to work well	for	recognizing	the	subjects
----------	----	------------	---	-----------	---	--------	---	---	---	---	---	---	---	---
questions
C inourdataset.		However,academicresultsoniNat21[4]
---------------	---	----------------------------------	---	---	---	---	---	---	---	---	---	---	---	---
andGLDv2[2]suggeststhatspecializedsystemsworkvery
								We replaced		the name	of the	category	C	in the ques-
---	---	---	---	---	---	---	---	-----------	---	--------	------	--------	---	------------
wellonthesedatasets,suggestingthatagoodimage-based tionbyitssupercategory. Thisrephrasingprocesswasrun
classifierwouldmakeaviablesubstituteforLens.
forallautomaticallygeneratedsingle-hopquestions(single
Generally, the core points of our paper are that (1) En- andmulti-answer). Forexample,thequestionWhereisthe
cyclopedic-VQA		poses a	hard challenge		for large	LLMs	E´glise
--------------	---	-------	--------------	---	---------	----	-------	------------	---	--------	---	-------------	---	-----------
								Saint-Cannat		located?		was rephrased		to Where is
and VLMs, and (2) solving it requires augmenting the thischurchlocated? Afterseveraliterations,wedecidedto
LLM/VLM with a retrieval component to access a knowl- add three examples in the prompt, using chain-of-thought
edgebase.	Theexactchoiceofretrievalcomponentisflex-
---------	-----------------------------------------	---	---	---	---	---	---------	---	----	------------	---	-----	---------	------
							reasoning		with	three steps.	The	model	generally	writes
ibleandlikelysubjecttofurtherexploration. the three required steps in the output, and we parse the
							rephrased		question	right	after the	third	step written	by the
---	---	---	---	---	---	---	---------	---	--------	-----	---------	-----	------------	------
D.Qualitativeresults model. Sometimes, the model would return an incorrectly
							rephrased		question.	We observe		that the	main	failure case
---	---	---	---	---	---	---	---------	---	---------	----------	---	--------	----	------------
We present qualitative results for some of the methods is when the model simply copies the input question to the
we experimented with in Figures 6 and 7. These illustrate output. Thus, we filter out any rephrased question that is
caseswherethedifferentsetupsmaywork,andcaseswhere
							identical		to the	original question		fed in	its input.	We sam-
---	---	---	---	---	---	---	---------	---	------	-----------------	---	------	----------	-------
allofthemfail. Forexample,givenatextualquestionWhat pled100rephrasedquestionstomanuallyassessthequality
does this reptile eat?, PaLM can reasonably guess an an- of the rephrasings, which we found to be correct in 90%
swer without any additional context: insects – since that’s ofthecases. ThefinalpromptisgiveninPrompt1,where
whatmanyreptileseat. “$C$” and “$Q$” are placeholders for the category C and
thetextual(pre-rephrasing)questionQ.
Formorespecificinformation,relatingtomoredetailed
properties(e.g.,numberofeggsaspecificreptilelays,dates,
							F.2.	Chaining		single-hop	questions		into	two-hop
---	---	---	---	---	---	---	----	--------	---	----------	---------	---	----	-------
howbigaspecificfishmaybecome),largemodelsmaystill
questions
makereasonableguesses,			butgenerallythisleadstoincor-
----------------------	------------------------------------	---	-----------------------------	---	---	---	---	----------	---	-----------------	---	-----------	---	-----------
rectanswers.	Augmentinglargemodelswithcontextfrom
								We created		two-hop questions		by chaining		two single-
retrieved knowledge improves the accuracy in these cases hopquestions,wheretheanswertothefirstsingle-hopques-
substantially,leadingtopreciseanswersthatareattributable tionservesasabridgeentitythatmakesaconnectiontothe
tothepieceofknowledgethatwasretrieved.
							second	single-hop		question.	For	example,	given	the first
---	---	---	---	---	---	---	------	----------	---	---------	---	--------	-----	---------
Finally, we also see cases where the generated answer single-hopquestion(SQ1)Whatisthemaincompetitorfor
is incorrect, even if the correct piece of knowledge is re- food for this animal? with answer (SA1) Spotted hyena,
trieved. Thisindicatesthatinsomecaseslargemodelsmay andthesecondsingle-hopquestion(SQ2)Whatisthepop-
stillhavedifficultiesusingretrievedknowledgetogenerate ulation size of this animal? with answer (SA2) Between
accurateanswers. 27,000 and 47,000 individuals, we generated What is the

Qualitative results: PaLI

	In which direction from the	What is Mary handling
In what month of 1944 was this What was this church fitted with
Question	german-czech border is this	St.Dominic in the facade of this
--------	----------------------------	---------------------------------	---------	---	-----------------	---	--------	---
					church destroyed?		in 1914?
	mountain located?		building?
No
	north		child		december		window
---	-----	---	-----	---	--------	---	------	---
Retrieval
Lens
	north		rosary		december		bell
---	-----	---	------	---	--------	---	----	---
Section
Oracle
	north		rosary		may		steeple
---	-----	---	------	---	---	---	-------	---
Section
Ground
	north		the rosary		may		electricity
---	-----	---	----------	---	---	---	-----------	---
Truth
Figure 6:
PaLIqualitativeresults. Wepresent4multi-modalquestionsandtheanswersproducedwithdifferentexperimentalsetups,
usingthePaLImodel: withoutretrieval(“NoRetrieval”), Lens-basedretrievalofKBsection(“LensSection”), OracleretrievalofKB
secQtionu(“aOrlaictleaSetcitvione”) .rAeddsitiounalltlys,w:e Fprolvaidenth-ePgroaunLd-tMruthanswerinthelastrow.Weshowcaseswhereeachofthesetupscan
producethecorrectanswer,aswellasanexamplewhereallmethodsfail(theright-mostone).
		How many eggs does this				How big does this fish typically
---	---	------------------------	---	---	---	---------------------------------	---	---
Question What does this reptile eat? What does this animal eat?
		reptile typically lay?					become?
---	---	----------------------	---	---	---	---	-------	---
No
	insects		10-15		insects		6 cm
---	-------	---	-----	---	-------	---	----	---
Retrieval
Lens
	insects		3-6		plankton		13 cm
---	-------	---	---	---	--------	---	-----	---
Section
Oracle	insects		3-6		mussels		13 cm
-------	-------	---	---	---	-------	---	-----	---
Section
Ground
	insects		3-6		mussels		25 cm
---	-------	---	---	---	-------	---	-----	---
Truth
Figure7: PaLMqualitativeresults. Wepresent4multi-modalquestionsandtheanswersproducedwithdifferentexperimentalsetups,
usingthePaLMmodel: withoutretrieval(“NoRetrieval”),Lens-basedretrievalofKBsection(“LensSection”),OracleretrievalofKB
section(“OracleSection”).Additionally,weprovidetheground-truthanswerinthelastrow.Weshowcaseswhereeachofthesetupscan
producethecorrectanswer,aswellasanexamplewhereallmethodsfail(theright-mostone).
population	size of the main competitor	for food	of this an-	individuals.
----------	---------------------------	--------	-----------	------------	---	---	---	---
imal? Inthisexample,theentityspottedhyenaservesasa
				After	several iterations,	we crafted	a prompt	with 4 ex-
---------------------------------------	---	---	-----------	-----	-------------------	----------	--------	----------
bridgebetweenthetwosingle-hopquestions.			Notethatthe
answertothetwo-hopquestionisidenticaltotheanswerto amplesofvaryingdifficulty(Prompt2). Theoutputofthe
				model	is taken as the chained	two-hop	question.	Some-
---	---	---	---	-----	-----------------------	-------	---------	-----
thesecondsingle-hopquestion,Between27,000and47,000
times,though,themodelwouldreturnanincorrecttwo-hop

NumberofQ+Apairs NumberofCategories %UniqueC Total(I,Q,A)triplets

			Train	Val	Test	Train	Val	Test	Val	Test	Train	Val	Test
Templated 8,133 200 500 1,983 114 300 26% 21% 40,665 1,000 500
12taNi Automatic 96,317 1,000 1,500 4,061 343 562 5% 6% 481,585 5,000 1,500
MultiAnswer 11,262 200 500 3,585 118 318 31% 22% 56,310 1,000 500
		TwoHop	10,819	200	500	2,499	131	359	11%	6%	54,095	1,000	500
---	---	---------	------	---	---	-----	---	---	---	---	------	-----	---	---
		Templated	5,795	200	500	1,808	135	268	27%	28%	25,870	827	500
2vDLG
Automatic 57,124 750 1,250 1,965 138 208 9% 9% 255,529 3,025 1,250
MultiAnswer 12,667 200 500 6,449 151 339 45% 56% 56,426 844 500
		TwoHop	10,221	200	500	1,405	99	154	11%	11%	45,771	895	500
---	---	------	------	---	---	-----	---	---	---	---	------	---	---	---
Total 212,338 2,950 5,750 16,249 1,071 2,152 14% 17% 1,016,251 13,591 5,750
Table8: DatasetStatistics. Wereportthenumberofquestion,answer,categoriesandtripletsforbothsupportingdatasetsanddifferent
questiontypes.Intotalourdatasetcontains1MVQAtriplets,makingitthelargestofitskind.
question. Forthisreason, wedesignedasecondpromptto evantsectionswithPaLM.Whennotavailable,werevertto
validatethetwo-hopquestionprovidedastheinitialoutput								usingPrompt4.
----------------------------------------------------	---	---	---	---	---	---	---	-------------	---	---	---	---	---	---
(Prompt3).Themodelisaskedtoanswerthetwo-hopques-
G.2.Promptforrelevantsectionidentification
tionusingthetwoinitialsingle-hopquestionswithanswers
ascontext.Ifthepredictedanswerisidenticaltotheanswer
								We use	Prompt	8 to	identify	relevant	sections	in a
---	---	---	---	---	---	---	---	------	------	----	--------	--------	--------	----
of the second single-hop question, then the two-hop ques- Wikipedia article retrieved by Lens. To do so, we query
tionisvalidatedandkeptinourdataset;otherwise,itisdis-
PaLMforeachsection$S$intheLensretrievedWikipedia
carded. Finally,wefiltersomecommonfailurecases,such article for a question $Q$. Note that we use a few exam-
aswhenthemodeloutputstheexactfirstsingle-hoporsec- plesintheprompttoconditionPaLMtogenerateayes/no
ond single-hop		question	as the	chained	two-hop	question.
--------------	---	--------	------	-------	-------	---------	---	-------	---------------------------------------	---	---	---	---	---
								answer.	TheanswerproducedbyPaLMisconvertedintoa
We sampled 100 chained two-hop questions to manually stringandmatchedtoeitheryesorno. Whennomatchis
assess their	quality,	which	we	found	to be correct	in 88%
------------	--------	-----	---	-----	-------------	------	---	---------	------	------------	------	--------	---	---------
								found, we	assume	that section	is not	relevant	to	the ques-
ofthecases. Inthefinalprompts(Prompts2-3)“$SQ1$”/ tion. Ifmorethanonesectionisidentifiedasrelevantfora
“$SA1$”areplaceholdersforthefirstsingle-hopquestion/ question, the input $S$ to Prompt 7 becomes the concate-
answer,“$SQ2$”/“$SA2$”areplaceholdersforthesecond
nationofallrelevantsections.
single-hopquestion/answer,and“$Q$”istheplaceholder
forthegeneratedtwo-hopquestion.
G.Promptsforevaluation
InSection5weanalyzetheperformanceoflargemodels
inourdataset.	Similarlytothedatasetcreationprocess,we
-------------	---------------------------------------	---	-------------------------	-----------	---	--------	---	---	---	---	---	---	---	---
use diverse	prompts	to	adapt	and improve	the	behavior	of
thesemodels.	Inparticular,		weusepromptstoincorporate
retrievalresultsandtoidentifyrelevantsectionsforarticles
retrievedwithLens.
G.1.Promptsforquestionanswering
We use	textual	Prompts	4-7	to produce		answers with
--------------	--------	-------------	--------	------------	------------	-------------	---	---	---	---	---	---	---	---
large models.	Note	that	we	use the	same prompts	when
using PaLI	or PaLM.		However,	we always		use the ques-
tion image	I as	an additional		input to	the model	when us-
ing PaLI.	Note	that we	use $Q$	to refer	to the	textual part
of a question,	$C$	to denote		the question	subject	C name,
$Art$todenotethefulltextofaWikipediaarticleand$S$
todenoteasectionofaWikipediaarticle.					Weusethesame
---------------------------------------	---	---	------------	-----	------------	-------------	---	---	---	---	---	---	---	---
retrievalpromptsforLensandOraclesetups.						However,for
Lens experiments,		we	only include	$Art$	or	$S$ if avail-
able,aswemightnotretrieveentitieswithLensorfindrel-

Prompt1:	Rephrasingautomatically-generatedsinglehopquestion
In this task, please rephrase the question by replacing the entity name by the word "this",
followed	by the type	of the	entity.	See the	examples	below:
--------	-----------	------	-------	-------	--------	------	---
EXAMPLE 1:
entity	name: eiffel	tower
---------	-------------	--------	------	------------	---	------	---------
question:	How tall	is the	eiffel	tower?
step 1	(find type of	entity):	The	eiffel tower	is	a type	of: tower
step 2 (write the word "this" followed by the type obtained in "step 1"): this tower
step 3	(final rephrased	question):		How tall	is this	tower?
------	----------------	----------	---	--------	-------	------	---
EXAMPLE 2:
entity	name: salmon
---------	-------------	--------	-----------	---------	------	-------	----
question:	Which country	is	the largest	producer	of	salmon?
step 1	(find type of	entity):	The	salmon is	a type	of:	fish
step 2 (write the word "this" followed by the type obtained in "step 1"): this fish
step 3 (final rephrased question): Which country is the largest producer of this fish?
EXAMPLE 3:
entity	name: Grand	Beach Provincial		Park
------	-----------	----------------	---	----	---	---	---
question: grand beach provincial park is located on the east side of what lake?
step 1 (find type of entity): The Grand Beach Provincial Park is a type of: park
step 2 (write the word "this" followed by the type obtained in "step 1"): this park
step 3 (final rephrased question): this park is located on the east side of what lake?
Note that the entity name should not be part of the rephrased question.
Please	make sure to	write out	all	3 steps as	in the	examples	above.
------	------------	---------	---	----------	------	--------	------
Based on the above examples, provide a rephrased question for the following case:
entity	name: $C$
---------	---------	---	---	---	---	---	---
question:	$Q$

Prompt2:	Chainingtwosingle-hopquestionsintoatwo-hopquestion
EXAMPLE 1:
question	1: in which	city	is this	building	located?
--------	----------------	------	-------	-----------	--------	----------	---
answer	1: San Francisco
question	2: what	is the	average	temperature	in San	Francisco?
answer	2: 15 Celsius
combined question: What is the average temperature in the city where this building is located?
EXAMPLE 2:
question	1: what	is the	predator	of this	animal?
--------	----------------	------	---------	-------	-----------	---	---
answer	1: Lion
question	2: What	is the	weight of	a lion	on average?
answer	2: 190 kilograms
combined question: What is the average weight of the predator of this animal?
EXAMPLE 3:
question	1: In what	country	is this	plant	found?
--------	-------------	----------------	--------	-----	---------	------------	---
answer	1: Australia
question	2: What	does australia’s		size	give it a	wide variety	of?
answer	2: landscapes	and	climates
combined question: What does the size of the country where this plant is found give it a wide
variety of?
EXAMPLE 4:
question	1: What	country	is this	plant	the national	flower	of?
--------	---------------	-------	-------	-----	------------	------	---
answer	1: South Africa
question 2: south africa is a member of the commonwealth of nations and what other
organization?
answer	2: the G20
------	----------	---	---	---	---	---	---
combined question: The country that this plant is the national flower of is a member of the
commonwealth	of nations	and	what	other	organization?
------------	----------	---	----	-----	-------------	---	---
Based on the above 4 examples, provide a combined question for the following case, such that
the answer to the combined question is the same as the answer to question 2:
question	1: $SQ1$
--------	------------------------------------------------------------	---	---	---	---	---	---
answer	1: $SA1$
question	2: $SQ2$
answer	2: $SA2$
combined	question:
Prompt3:	Validatingtwo-hopquestiongiventheoriginalsingle-hopquestions
question	1: $SQ1$
answer	1: $SA1$
question	2: $SQ2$
answer	2: $SA2$
Based on the questions and answers above, please answer the following question: $Q$

Prompt4:	Questiononlyevaluation
Question:	$Q$
The answer	is:
Prompt5:	Retrievalwithentityname
Entity	name:	$C$
Question:	$Q$
The answer	is:
Prompt6:	RetrievalwithKBArticle
Context:	$Art$
Question:	$Q$
The answer	is:
Prompt7:	RetrievalwithKBSection
Context:	$S$
Question:	$Q$
The answer	is:
Prompt8:	LensretrievaltogetKBSection
Can the	answer	to the	question	be found	in the text?
Question:	Is	this	fungus edible?
Text: This compound induces mammalian cells (specifically, the cell line HL60 to differentiate
into granulocyte- or macrophage-like cells. The fungus also contains the mycotoxin muscarine,
and the antifungal metabolite strobilurin D. Despite the presence of these toxins, some guides
list this	fungus	safe	for human	consumption.
----------	------	------	---------	---------------	-------------	---
The answer	is:	yes
Can the	answer	to the	question	be found	in the text?
Question:	In	which	season	does this plant	give flowers?
Text: This cactus has stems about 1/2-1 inch wide with 6-9 edges. Its flowers are white, up
to 30 centimetres in diameter with a scent redolent of vanilla. The flowers open after sundown,
closing and wasting after a few hours. By 9 am the next day they are gone.
The answer	is:	no
----------	------	------	-----------	--------	------------	---
Can the	answer	to the	question	be found	in the text?
Question:	What	is	the habitat	of this	animal?
Text: X is native to Europe and North Africa through to Central Asia. It is introduced to the
United States and parts of South America. It widespread across the northeastern United States
and eastern Canada, and can be found outside, or more commonly inside houses. It is thought to
have been	introduced		into America	from	Europe by English	colonists.
----------	----------	------	------------	--------	-----------------	----------
The answer	is:	yes
Can the	answer	to the	question	be found	in the text?
Question:	$Q$
Text: $S$
The answer	is:
----------	---	---	---	---	---	---