q116/random_k4/random_3.md · luoojason/muddle-eval-bundle-md at main

		Learning		Fine-Grained		Grounded	Citations		for
			Attributed		Large	Language	Models
LeiHuang1,XiaochengFeng1,2*,WeitaoMa1,YuxuanGu1,WeihongZhong1,XiachongFeng1
WeijiangYu3,WeihuaPeng3,DuyuTang3,DandanTu3,BingQin1,2
1HarbinInstituteofTechnology,Harbin,China
		2 PengChengLaboratory				3HuaweiInc.,Shenzhen,China
---	---	---------------------	---	---	---	--------------------------	---	---	---	---	---
{lhuang,xcfeng,wtma,yxgu,whzhong,xiachongfeng,qinb}@ir.hit.edu.cn
{weijiangyu8,pengwh.hit}@gmail.com,{tangduyu,tudandan}@huawei.com
Abstract
Why is it bad to eat cookie dough for risk of salmonella but things
like Cookie Dough Bites are ok?
Despite	the	impressive	performance		on
-------	---	----------	-----------	---	---	--------------------------------------------------------	---	---	---	---	---
						Salmonella is most commonly caused by eating undercooked					1
information-seekingtasks,largelanguagemod- or raw foods like eggs or meat. (...) You know how your mom
4202 guA 8 ]LC.sc[ 1v86540.8042:viXra els (LLMs) still struggle with hallucinations. always warned you not to eat raw cookie dough? This is why.
Attributed	LLMs,	which	augment	generated
----------	-----	-----	-------	---------	---	---------------------------------------------------------------	---	---	---	---	---
						Food safety recommendations encourage people to cook eggs until					2
text within-line citations, haveshownpoten- the white and yolk are firm in order to kill any bacteria. (...) However,
tial in mitigating hallucinations and improv- anyone making cookies can do things to reduce this risk by
using pasteurized egg products.
ingverifiability.		However,currentapproaches
-----------------	---	-------------------------	---	---	---	---	---	---	---	---	---
suffer from suboptimal citation quality due Raw cookie dough can carry the risk of salmonella contamination, which can
to their reliance on in-context learning. Fur- lead to foodborne illness if the dough is eaten raw 1 . However, products
like Cookie Dough Bites are generally safe because they are made with
thermore,	the	practice	of citing	only	coarse	pasteurized eggs 1 2 .
---------	---	--------	---------	----	------	---------------------------	---	---	---	---	---
documentidentifiersmakesitchallengingfor
Difficult to verify and subject to
userstoperformfine-grainedverification. In hallucinations and citation errors
this work,	we	introduce	FRONT,	a	training
----------	---	---------	------	---	--------	---	---	---	---	---	---
G R O U N D I N G 1 salmonella is most commonly caused by eating
framework designed to teach LLMs to gen- undercooked or raw foods like eggs or meat 2 Anyone making cookies
can do things to reduce this risk by using pasteurized egg products.
erateFine-gRainedgrOuNdedciTations. By A N S W E R Eating raw cookie dough can be risky due to salmonella,
groundingmodeloutputsinfine-grainedsup- commonly found in raw eggs 1 . However, this concern is mitigated in
products like Cookie Dough Bites, as the products use pasteurized egg
portingquotes,thesequotesguidethegenera-
products, which are safe to eat raw 2 .
tionofgroundedandconsistentresponses,not
onlyimprovingcitationqualitybutalsofacili- Interpretable and accurately cited
tatingfine-grainedverification.Experimentson
						Figure1:	Comparedwiththecurrentattributedsystems,
---	---	---	---	---	---	--------	----------------------------------------	---	---	---	---
theALCEbenchmarkdemonstratetheefficacy
thecoreideabehindFRONTistofirstselectthesupport-
ofFRONTingeneratingsuperiorgroundedre-
						ing quotes	from	retrieved	sources and	then	condition
------------------------------------	---	---	---	---	----	----------	----	---------	-----------	----	---------
sponsesandhighlysupportivecitations.					With
thegenerationprocessonthem,ensuringgroundedre-
LLaMA-2-7B,theframeworksignificantlyout-
sponsesandaccuratecitations.
performsallthebaselines,			achievinganaver-
------------------------	---	---	----------------	---	---	---	---	---	---	---	---
ageof14.21%improvementincitationquality
acrossalldatasets,evensurpassingChatGPT1. SuchprevalenceofhallucinationsinLLMout-
						puts has	motivated	the	development	of	attributed
---	---	---	---	---	---	--------	---------	---	-----------	---	----------
1 Introduction
						systems	(Nakano	et al.,	2021; Thoppilan		et al.,
---	---	---	---	---	---	-------	-------	-------	---------------	---	-------
Therecentadventoflargelanguagemodels(LLMs) 2022;Menicketal.,2022),suchasNewBing2 and
(Touvronetal.,2023;OpenAI,2023;Zhaoetal., Perplexity3,whereLLMsareallowedtogenerate
2023) has taken the world by storm, fueling a responses with in-line citations. Not only does it
paradigm	shift in	information		acquisition	(Zhu
--------	--------	-----------	---	-----------	----	---	---	---	---	---	---
improvefactualityandalleviatehallucinations,but
etal.,2023). Despitetheircompellingperformance, italsosimplifiesuserverificationofmodeloutputs,
LLMs still struggle with hallucinations (Ji et al., furtherenhancingtheverifiabilityofLLMs.
2023;Huangetal.,2023),atendencytofabricate Despiterecentadvancements,currentattributed
non-existent	facts	or generate	unfaithful		content.
------------	-----	-----------	----------	---	--------	--------------------------------------	---	---	---	---	--------
						LLMsstillexposesignificantlimitations.					Firstly,
This issue further poses a risk of misinformation current approaches predominantly rely on either
dissemination(ChenandShu,2023),directlyim- in-contextlearning(Gaoetal.,2023b)orpost-hoc
pactingthereliabilityandtrustworthinessofLLMs.
retrieval(Gaoetal.,2023a)toachieveattribution,
*CorrespondingAuthor
1Our data and code can be found at: https://github. 2https://www.bing.com/chat
com/LuckyyySTA/Fine-grained-Attribution. 3https://www.perplexity.ai

							2 RelatedWork
lacking	an inherent	attribution		capability		within
LLMs,therebyresultingincompromisedcitation
							RetrievalAugmentedGeneration.					Recently,re-
-----------------------	---------	---	-----------------------	----	---------	----	-----------------------------	---	----------	---	-----	------------	---
quality(Liuetal.,2023).			Secondly,thesecitations
							trieval augmented		generation		(RAG)	(Karpukhin
are typically	presented		in the	form	of either	doc-
ument identifiers (Nakano et al., 2021) or URLs etal.,2020;Lewisetal.,2020;Fengetal.,2023;
Gaoetal.,2023c)hasshownpromiseinknowledge-
(Thoppilan	et al.,	2022).	Such	coarse	attribution
----------	-------	------	----	------	-----------	---	---------	------	----------------	---	---------	---	-----
							intensive	tasks.	By incorporating		retrieved		docu-
complicatestheprocessforuserstoperformfine-
							ments,	LLMsareequippedwithup-to-dateinfor-
---	---	---	---	---	---	---	------	-----------------------------------	---	---	---	---	---
grainedverification(e.g.,pinpointspecificsupport-
							mation,	significantly	mitigating		knowledge		gaps.
---	---	---	---	---	---	---	-------	-------------	----------	---	---------	---	-----
ingevidence),particularlyinlengthydocuments.
							However,	recent	studies	(Shi	et al.,	2023;	Yoran
-------	----	------	----------	---	----------	----	--------	------	-------	----	-------	-----	-----
To this	end,	we aim	to advance		attributed	text
etal.,2023;Xuetal.,2023a;Zhuetal.,2024)have
generationbyempoweringLLMswithfine-grained
attributionability. However,onechallengecomes revealed that existing retrieval-augmented LLMs
							struggle	to handle	irrelevant		or contradictory		re-
---	---	---	---	---	---	---	--------	---------	----------	---	----------------	---	---
fromtheacquisitionofhigh-qualityattributiondata
trievaldocumentsandeffectivelyutilizecontextual
for supervised		fine-tuning,		which	is difficult	and
------------------------------------	---	------------	---	-----	------------	-------	---------	-----	-----------	---	----------	---	-------
							evidence.	These	limitations		can result	in	perfor-
costlytoannotate,andthereforescarce.						Thus,we
mancedegradationorevenhallucinations(Huang
start with	an automatic		data	generation		pipeline
----------	------------	---	----	----------	---	--------	--------------	---	------------	---	---------	---	----
							et al., 2023),		highlighting	the	necessity	for	more
tailoredforcollectinghigh-qualityattributiondata
factualandverifiablesystems.
(§3.1). Givenauserquery,thepipelineautomates
dataconstructionthroughdocumentretrieval,rele-
							AttributedLargeLanguageModels.					Theper-
---	---	---	---	---	---	---	------------------------------	---	---	---	---	-------	---
vancereranking,attributedanswergeneration,and
							sistent	challenge	of hallucinations			within	LLMs
-------------------------------	---	------	----	-------------------	---	---	-----------	---------	-----------------	---	-------------	------	----
data filtering	to	ensure	both	the informativeness
							has spurred	the	development		of attributed		LLMs
andattributabilityoftheanswers.				Furthermore,to
(Bohnetetal.,2022;Lietal.,2023;Worledgeetal.,
betterunlockLLMs’abilityforfine-grainedattri-
2023),whichseektoenhanceinformationverifia-
bution,	we introduce		FRONT,	a two-stage		train-
-------------	------------	----	-------	-----------	---	--------	---------	----------	---------	------	---------	-----------	--------
							bility by	generating	responses		with	attribution	to
ing framework		that	teaches	LLMs	to	generate
							evidence	sources.	The	way of	providing		attribu-
Fine-gRainedgrOuNdedciTations(§3.2),consist-
							tionsvariesacrossstudies.			Forexample,Gaoetal.
------	---------	------	----------	---	----	---	-------------------------	---	---	-------------------	---	---	---
ing of	Grounding	Guided	Generation		(G3)	and
(2023b)enablesLLMstogeneratetextwithin-line
Consistency-AwareAlignment(CAA).Duringthe
							citations	via in-context		learning.	Another		line of
---	---	---	---	---	---	---	---------	--------------	---	---------	-------	---	-------
G3
stage,theLLMfirstselectssupportingquotes
							research	(Gao	et al., 2023a;	Xu	et al.,	2023b)	ex-
---	---	---	---	---	---	---	--------	----	--------------	---	-------	------	---
fromretrievedsources(grounding)andthencondi-
plorespost-hocattribution,whereLLMsfirstgen-
tionsthegenerationprocessonthem(generation).
erateaninitialresponseandthenretrievethemost
TheCAAstagethenutilizespreferenceoptimiza-
							relevantevidencetoachieveattribution.					Inthispa-
---	---	---	---	---	---	---	-------------------------------------	---	---	---	---	---------	---
tiontofurtheralignthegroundingandgeneration
per,weadvancetheresearchonattributedLLMs
processbyautomaticallyconstructingpreference
							further.	Unlikeexistingmodelsthatpredominantly
---	---	---	---	---	---	---	--------	-------------------------------------	---	---	---	---	---
signals. Inthisway,thesequotescanserveasfine-
							cite document		identifiers,	we	delve	into	a more
---	---	---	---	---	---	---	-------------	---	------------	---	-----	----	------
grainedcitationsandimprovetheefficiencyofthe
fine-grainedformofattributionbypinpointingand
verificationprocessforusers(seeFigure1).
citingspecificextractivequotes.
We conduct		extensive	experiments		to	evaluate
----------	---	---------	-----------	---	---	--------	---	---	---	---	---	---	---
our framework on the ALCE Benchmark (Gao 3 TaskFormulationandMethodology
etal.,2023b).	Ourfindingsareasfollows:
-------------	------------------------	---	---	---	---	---	---	---	---	---	---	---	---
Following(Liuetal.,2023;Gaoetal.,2023b),the
• FRONT	demonstratessupervisorperformance
-------	---------------------------------	---	---	---	---	---	------------------	---	-----------	---	-----	------	-----
							task is formalized		as follows:		given	a user	query
gainsincitationqualitycomparedtoallbase- q and a corpus of retrieved documents D as in-
lines,achievinganaverage14.21%improve-
							put, the	LLM	is required	to	produce	a response
---	---	---	---	---	---	---	--------	---	-----------	---	-------	----------	---
mentusingLLaMA-2-7B.
							S, which	consists	of statements		with	embedded
---	---	---	---	---	---	---	-----------------	--------	--------------------	---	----	--------	----
							in-linecitations.		WeassumetheresponseS				com-
• Humanevaluationrevealsthatthequotesgen-
							prising	with	n statements	S	= {s	,s ,...,s	}
----------------------------------------	---	-------------	---	---	-------	-------	-----------------	--------	------------	-----------------------	----	-------------	---
erated	by	our framework		are	of high	quality						1 2	n
							andeachstatements			∈ S,citesalistofpassage
andsignificantlybenefituserverification.									i
							C i = {c	i1 ,c i2	,...}, where	c ij	∈ D.	Specifically,
• AnalysisshowsthatFRONTgenerateslesshal- citationsarepresentedintheformof[1][2].
lucinationanddemonstratesremarkablegen- Next,wepresentacomprehensiveoverviewof
eralizationacrossdifferentbasemodels. ourmethod,whichconsistsoftwoprimarycompo-

						Demonstrations			1 3 . Another suggests (...) and transport 2 .
					Instruction: Generate an answer and cite the				Informative?
Retrieve source for the question and provided documents Break into
Question: Why do bagels have holes in the middle？
	Top-100 retrieval documents										statements
---	---------------------------	---	---	---	---	---	---	---	---	---	----------	---
(...)
					Prior to (...) the hole was originally put in place				Statement 1	Bagel holes ensures even baking by
---	---	---	---	---	---------------------------------------------------	---	---	---	-----------	----------------------------------	---	---
allowing heat to circulate around a
					to allow for easier handing prior.					greater surface area 1 3 .
---	--------	--------------------------------	----------------------------------	---	----------------------------------	---	---	---	---	--------------------------------	---	---
	T h e b	a k e r s (. .. ) th e h o	l e i n t h e c e n tre of the
1 ba g e l e n s u r e s t h a t i t b a k e s e v e n ly . Statement 2 A n o t h e r s u g g e s t s t h a t t h e y w e re u s ed
										fo r e a s i e r	h a n d li n g a n	d tr a n sp o rt 2 .
------	---------------------------------------------	---	---	---	------------------------------------------------	---	---	---	-----------	------------------	--------------------	---------------------------
	Prior to (...) the hole was originally put in				Bagel holes ensures even baking by allowing
	2				heat to circulate around a greater surface							Support
	place to allow for easier handing prior.								Attributed?
Rerank					area 1 3 . Another suggests that they were
T h e r i n g s h a p e a l lo w s h e a t t o c ir c u la t e used for easier handling and transport 2 . Documents
	3 ar o u n	d ( .. . ) m a k e s b a g	e l s t o c o o k fa s t er
---	----------	------------------------------	---------------------------------	---	---	---	---	---	---	---	---	---
Figure2: Overviewofthedatagenerationpipeline. Thepipelineconsistsofthreeprimarysteps: datacollection,
answergeneration,anddatafiltering. Firstly,givenauserquery,thedatacollectionmoduleretrievesthetop100
relevantdocumentsandemploysarerankingmodeltoselectthetop5mostpertinentdocuments. Subsequently,
attributedresponsesaregeneratedbydistillingChatGPTviain-contextlearning. Finally,allresponsesarefiltered
bythedatafilteringmoduletoensureinformativenessandattributability.
nents: anautomaticdatagenerationpipeline(§3.1) subsequently re-rankedby RankVicuna (Pradeep
andatwo-stagetrainingframework(§3.2). etal.,2023)consideringitssuperiorperformance
						inlistwisere-ranking,			resultinginthetop5most
---	---	---	---	---	---	---------------------	---	---	----------------------	---	---	---
3.1 AutomaticDataGenerationPipeline
relevantdocumentsforeachquery.
EquippingLLMswiththeattributioncapabilityne-
						Attributed	Answer		Generation.		Given	the re-
----------	--------	----	-------------	------------	--------	----------	-----------	---	-----------	-------	-----	----------
cessitates	training	data	that includes	high-quality
						markable	performance		of	ChatGPT	in	attributed
responses	paired	with	precise	citations,	which is
questionanswering,weemployChatGPTtogener-
typically	labor-intensive		and	costly. To	address
---------	---------------	---	---	----------	-------	---	---	---	---	---	---	---
ateanswerswithcorrespondingcitationsforgiven
thischallenge,weproposeapipelinedesignedfor
						queriesandthetop5retrieveddocuments.						Wepro-
---	---	---	---	---	---	------------------------------------	---	---	---	---	---	------
theautomaticgenerationofhigh-qualityattributed
videpreciseinstructionsandin-contextdemonstra-
data4.
Thispipelinecomprisesthreecorecompo-
tionstoensurethatChatGPTproducesinformative
nents: datacollection,attributedanswergeneration,
responsesandcitesthesourcesaccordingly.
anddatafiltering,asoutlinedinFigure2.
						DataFiltering.		Toguaranteethehighqualityof
---------------	---	----------------------------	---	---	---	--------------	---	---------------------------	-----	---------	---	-----------
DataCollection.		Tosimulatethereal-worldenvi-
						our synthetic		training	data,	we employ		a data fil-
ronmentforinformation-seeking,wecollectques-
teringprocessguidedbytwokeycriteriaderived
tionsfromtheAQuAMuSedataset(Kulkarnietal.,
						fromKamallooetal.(2023):				(1)informativeness:
---	---	---	---	---	---	------------------------	---	---	---	-------------------	---	---
2020),whichisderivedfromtheNaturalQuestion
						assessing	if the	answer	provides		sufficient	infor-
-----------------------------------	---	---	---	---	-----	------------------------------------------	------	------	--------	---	----------	------
(NQ)dataset(Kwiatkowskietal.,2019).					TheNQ
						mationtothequestion,and(2)attributability:						de-
datasetcomprisesrealuserqueriesfromtheGoogle
						termining	if the	answer	is	attributed		to the cited
---	---	---	---	---	---	---------	------	------	---	----------	---	------------
searchengine,providingarobustbasisforrealistic
						documents.	To	mitigate		the impact	of	nonsensi-
------------------	---	----------	---	-----------	-----	----------	---	--------	---	----------	---	---------
question-answering		scenarios.		The dataset	spans
calqueriesandirrelevantdocumentretrievalthat
a range	of diverse	question	types,	demanding	an-
-------	----------	--------	------	---------	---	--------	------------------	---	---	--------	---	----------
						may lead	to non-informative			answers,		we utilize
swersofvaryinglengths,fromconcisetodetailed.
ChatGPTforpreliminaryinformativenessannota-
Tomimicthewayasearchenginemightsynthesize
						tions. Responses		categorized		as	non-informative
---	---	---	---	---	---	----------------	---	-----------	---	---	---------------	---
documentsofhighrelevanceinresponsetoauser
						aredirectlyexcluded.			Furthermore,toensurethat
------	---------	------	-------	-------	--------	--------------------	---	---	------------------------	---	---	---
query,	we employ	Sphere	(Piktus	et al.,	2021), a
answersareaccompaniedbyhighlysupportiveci-
pre-processedandcleanedversionoftheCommon
tations,wetrainadiscriminatoronhuman-labeled
Crawlcorpus,servingasaproxywebsearchindex.
						data from	the	comprehensive		evaluation		by Liu
---	---	---	---	---	---	---------	---	-------------	---	----------	---	------
Inparticular,foragivenuserquerysampledfrom
						et al. (2023),		where	attributability		is	categorized
---	------------	----------	------------	----------	----	----------------	-------------------------------------	-------------------------------	---------------	-----	---	-----------
the	AQuAMuSe	dataset,	we initially	retrieve	the
						intothreelevels:		fullsupport,partialsupport,orno
top	100 relevant	documents	from	the Sphere	cor-
						support.	Wequantitativelymapthediscriminator’s
pus	using sparse	retrieval.	These	documents	are
						outputs	to an	attributability		score	and	ultimately
4Attributeddatarefersto“answerswithin-linecitations”. deriveanaveragescoreforeachattributedanswer.

Q : W h y d o	b a g e l s h a v e
		G r o u n d i ng	T h e b a k e r s ( . .	. ) t h e h o l e	i n t h e c e n t r e o f the	Q	u e s t i o n : W h y	d o b a g e l s	G R D eI G 1l e	n s u r e s t h a t i t
h o l e s i n t h e m i d d l e ？ 1 b a g e l { e n s u r e s t h a t i t b a k e s e v e n l y } . h a v e h o l e s i n t h e m i d d l e ？ b a kO eU Ns N v e n y . 3 a l l o w s h e a t
							( . . . )		t o c i r c u l a t e a r	o u n d . 2 a l l o w s A u N r S f W a c E	R e a B r a e g a e 1 l h 3 o A l e n s o ( t . h . . e ) r	g ( r . e .. a ) t
---	---	---	---	---	---	---	---------	---	-----------------------------	-------------------------------------------------------------	-----------------------------------------------------------------------------------	--------------------
B a g e l h o l e s e n s u r e s e v e n b a k i n g b y P r i o r t o ( . . . ) t h e h o l e w a s o r i g i n a l l y p u t in P r i o r t o ( .. . ) p l a c e t o a ll o w f o r e a s i e r h a n d i n g p r i o r . h a n d l i n g a n d t r a n s p o r t 2 .
al lo w i n g h e a	t to c i r c u l a t e	a r o u n d a 2
--------------------	--------------------------	-------------------	---	---	---	---	---	---	---	---	---	---
greater surface area 1 3 .Another p l a c e t o { a l l o w f o r e a s ie r h a n d i n g p r i o r } . fo r e a s i e r h a n d i n g p r i o r . Consistency
suggests that they were used for		3	The ring shape {allows heat to circulate
--------------------------------	---	---	----------------------------------------	---	---	---	---	---	---	---	---	---
easier handling and transport 2 . around} (...) makes bagels to cook faster Gold Grounding Consistent Answer Contrastive
Large Language
	Supervised Fine-tuning						Model
--------	----------------------	--------------	-------------	---------------------------------------------	-------------------------------------------------	---	-----	---	--------------	-------------------	---	---
									Gold Grounding	Inconsistent Answer
Question			G R	Ol Uo N D sI N Gh 1 e n s	u r e s t h a t it b a k e s e v e n ly.
		Large Language	f3 a	l w e a t t o	c i r c u l a t e a r o u n d . 2 a l lo w
			o r	e a s i e r h a n d i n	g p r i o r .
Documents Model Direct Preference Optimization (DPO) A N S W E R B a g e l h o l e s (. . . ) f o r
			A A N n S o	W t h E e R r B (. a .. g ) e h l a h n o d le	lin s g ( . a .. n ) d g r t e ra a n t s s p u o rf r a t c 2 e . area13					a e s t h e	t i c s a n d u n i q u e t e x	t u r e 1
---	---	---	----------------	------------------------------------------------------------	-------------------------------------------------------------------------------------	---	---	---	---	-----------------	-------------------------------------	-------------
										( . . . ) h a	n d l i n g a n d tr a n s p o	r t 3 .
Figure 3: Overview of FRONT: The training recipe consists of two stages: grounding-guided generation and
consistency-aware alignment. It enables LLMs to first generate precise grounding and subsequently guide the
generationofattributedanswers,therebyenhancingfine-grainedattributioncapability.
Answersfallingbelowadefinedthresholdaresys- fine-tunedtogeneratearesponseS whichconsists
tematicallyexcludedtoensurethesyntheticdata’s of two components: the grounded quotes G and
reliability, which results in nearly 8,000 entries. theattributedanswerA. Specifically,thegrounded
Formoredetails,pleaserefertoAppendixA. quotesG aredelineatedasfollows:
3.2 Two-StageTrainingRecipe G = {[GROUNDING],(i ,e ),...,(i ,e )}, (1)
										1 1	n n
-------------------------	---	---	----------------	---	---	---	---	---	---	---	---	---
Inthissection,weintroduce			FRONT,atwo-stage
where[GROUNDING]denotesaspecialtokenindi-
trainingframeworkthataimsatempoweringLLMs
						cating	the	start of	the grounding	process.		Each
-----------------	---	-----------	-----------	---	--------	------	---	--------	-------------	--------	---	----
with fine-grained		attribution	capability.		Figure 3
tuplewithinG,comprisingadocumentidentifieri
illustratestheoverviewofourframework.
andthecorrespondingextractivesegmente,collec-
tivelyformingagroundedquote.
3.2.1 GroundingGuidedGeneration
						Similarly,		the formulation		of the	attributed	an-
----------	---	---------	------------	---	--------	----------	---	---------------	---	------	----------	---
To empower		LLMs with	fine-grained		attribu-
swerAisconciselypresentedas:
tion capability,		we propose	Grounding		Guided
----------------	---	----------	---------	---	------	---	---	---	---	---	---	---
(G3),
Generation which teaches LLMs to gener- A = {[ANSWER],s ,s ...,s }, (2)
										1 2	m
-------------------------	---	---	------------------	---	---	---	---	---	---	---	---	---
atefine-grainedcitations.			ThecornerstoneofG3
liesinenablingLLMstoextractsupportingquotes where [ANSWER] is a special token that signals
fromthesourcedocuments,eachassociatedwithits the beginning of the answer generation process.
documentidentifier,whichinturnguidesthegen- Each statement s cites a list of passages C =
								i				i
---------------------------	---	---------------------------	-----------------	---	---	-----------------	------------	------	---	------------------	----	-----
erationofattributedanswers.			Suchaformatoffers
						{c i1	,c i2 ,...},	wherec	ij	⊆ {i 1 ,i 2 ,...,i	n },	asde-
twoprimarybenefits.		Firstly,thedirectextraction				finedinEquation2.
of quotes from sources significantly reduces the Thus,thetraininglossisformulatedas:
impactoftheincorporationofirrelevantinforma-
tion and	the risk	of hallucinations		in	subsequent			N
--------	--------	-----------------	---	---	----------	---	---	---	---	---	---	---
(cid:88)
								L = −	logP(y		q ,D	;θ)
------------------	---	----------------------------	---	---	---	---	---	-----	------	-----	---	---
attributedanswers.		Secondly,theprocessnaturally								i i	i
i=1
facilitatesaccurateattribution,witheachdocument
identifierservingasaclearsupervisedsignalthat
						where	y	represents	the	combined	output	of
---	---	---	---	---	---	-----	---	----------	---	--------	------	---
i
delineatestheoriginoftheextractivequotes,thus
groundedquotesGandtheanswerAforeachgiven
improvingthecitationquality. queryq andsetofretrieveddocumentsD .
							i				i
---	---	---	---	---	---	---	---	---	---	---	---	---
However,theabsenceofspecificgroundingcon-
tent for statements within our generated dataset 3.2.2 Consistency-AwareAlignment
posesadditionalchallenges. Totacklethis,weem- While G3 unlocks the ability to first extract sup-
ploy ChatGPT to meticulously extract segments portingquotesbeforegeneratingattributedanswers,
fromciteddocumentsthatsupportthecorrespond- it occasionally leads to inconsistencies between
ingstatement. Hence,whengivenaqueryqandthe groundedquotesandattributedanswers. Suchdis-
top-5retrieveddocumentsD asinput,theLLMis crepancieschallengetheattempttoemploythese

	The cornerstone		of	our approach	involves	con-
4 ExperimentalSettings
trasting		a consistent		answer	with an	inconsis-
---------------	---	------------	--------------------------	--------	-----------	---------	------------	---	---	---	---	---	---
tent	one	under	the	guidance	of the same	oracle	4.1 Datasets
groundedquotes.			Thisalignswiththeconceptof
WeconductexperimentsontheALCEbenchmark
Reinforcement			Learning	from	Human	Feedback
-------------	---	---	--------	----	-----	--------	---	---	---	---	---	---	---
(Gaoetal.,2023b),designedforattributedtextgen-
(RLHF)		(Ouyang	et	al., 2022;	Bai et	al., 2022),
------	---	-------	---	----------	------	-----------	---	---	---	---	---	---	---
eration. Thebenchmarkincludesthreelong-form
whereLLMsarefurtherfine-tunedtodistinguish
QAdatasetsthatspanvarioustypesofquestions.
betweendesirableandundesirableresponsesunder
preference feedback. However, such contrastive ASQA (Stelmakh et al., 2022) is a long-form
preferencefeedbacktypicallycomesfromhuman factoidQAdatasetcharacterizedbyinherentlyam-
annotation.		Toautomaticallyconstructpreference
-----------	---	----------------------------------	---	---	---	---	-------	---------	----	-------	--------	-----	---
							biguous	questions	that	require	multiple	short	an-
pairsforpreferenceoptimization,weutilizetheat- swerstoencapsulatedifferentviewpoints.
tributedanswersgeneratedbysmallerLLMs(e.g.,
LLaMA-2-7B)underthein-contextlearningsetting ELI5 (Fanetal.,2019)featuresopen-endedques-
tionsintendedforsimplificationtothecomprehen-
asnegativesamples.				Theseanswers,characterized
------------------	---	---	---	--------------------------	---	---	---	---	---	---	---	---	---
sionleveloffive-year-olds,requiringexplanatory
bytheirlowqualityandinconsistencywithoracle
multi-sentenceresponses.
groundedquotes,automaticallyserveascontrastive
supervisionsignalswhenpairedwithhigh-quality
							QAMPARI		(Amouyaletal.,2022)isafactoid
-------------------------------	---	---	---	---	---------------	---	-------	---	-----------------------------	---	---	---	---
attributedanswerslabeledin§3.1.					Inthisscenario,
QAdatasetderivedfromWikipedia,whereanswers
theprocessnotonlyencouragestheLLMtogener- arestructuredasacompilationofentities.
ateattributedanswersthataremoreconsistentwith
thegroundedquotesbutalsofacilitatestheidentifi- 4.2 EvaluationMetrics
cationandcorrectionofnuancederrorspresentin Following the ALCE benchmark (Gao et al.,
smallermodels. 2023b), our evaluation primarily focuses on two
	Specifically,		we adopt	Direct	Preference	Opti-
---	-------------	---	--------	------	----------	-----	---------------	---	--------	---	-------	------------	---
							key dimensions:		Citation		Quality	and Correct-
mization(Rafailovetal.,2023),avariantofRLHF ness. Detaileddescriptionsofadditionalevaluation
known for its stability, for our contrastive align- dimensionsarepresentedintheAppendixB.
ment. Formally,foreachinstance,giventheoracle
							CitationQuality.		Citationqualityiscriticalfor
---	---	---	---	---	---	---	----------------	---	----------------------------	---	---	---	---
groundedg(i)alongwithaconsistentoracleanswer
(i)					(i)		evaluating	LLM	attribution,		assessed	along	two
---	---------------------------	---	---	---	---	---------	-----------	---	------------	-------	-----------	-----	---
y	aswellasanattributedanswery					generated
w					l
							dimensions:	(1)	Citation	Recall,	determining		if
byaweakerLLMviain-contextlearning,wecan
							the output	is entirely		supported	by	the cited	doc-
---	---	---	---	---	---	---	----------	-----------	---	---------	---	---------	----
simplyconstructapreferencedataset:
							uments,	and (2)	Citation	Precision,		assessing	if
---	---	---	---------------------	---	-------------	---	----------------------------------------------	-------	--------	----------	---	---------	---
		D	= (cid:8) x(i),τ(i),τ		(i)(cid:9)N ,
				w		(4)	eachcitationsupportsitscorrespondingstatement.
l i=1
	(i)		(i)				EvaluationisconductedbyTRUE(Honovichetal.,
------	---	--------	---	-------------------------	---	---	------------------------------------------	---	---	---	---	---	---
whereτ		= g(i)◦y		denotestheconcatenationof
	w		w				2022),aT5-11Bmodelfine-tunedonacollection
theoraclegroundingwiththeconsistent,attributed
ofNLIdatasetstoautomaticallyexaminetheentail-
		(i)	g(i)◦y	(i)
--------	---	------------	------	-----------------------	-------	-------	------------------------------------------	---	-------	----------	-------	---	------
answer,τ		=		denotestheconcatenation
		l		l			mentofciteddocumentsandthemodelgeneration.
with	the	inconsistent		attributed	answer.	Here, ◦
							Additionally,	to	capture	a holistic	measure		of ci-
signifiestheoperationofstringconcatenation.
tationquality,wealsoreporttheCitationF1,the
	Finally,wecanoptimizethepolicymodelπ					on
---	------------------------------------	---	---	---	---	---	-----------------------------------------	---	---	---	---	---	---
						θ	harmonicmeanofcitationprecisionandrecall:
thedatasetD
byminimizingthefollowingloss:
citationprecision·citationrecall
	L	(π ;π	;D)
---	---	-----	---	---	---	---	--------	---	---	---	---	---	-----
	DPO	θ	ref				F 1 = 2·						, (6)
citationprecision+citationrecall
			(cid:20)	(cid:18)	π (τ		x)
---	---	---	--------	--------	----	---	---	---	---	---	---	---	---
	=−E				θ	w
				logσ	βlog
(x,τw,τ l )∼D π (τ	x) (5) Correctness. FortheASQAdataset,correctness
ref w
π (τ	x) (cid:19)(cid:21) is quantified using exact match recall (EM Rec.)
θ l
	−βlog			,
---	-----	---	------	---	---	---	-----------	-------	---	---------	-------	---	-----
		π	(τ	x)				by checking	whether		the short	answers	are
			ref l

			ASQA			ELI5		QAMPARI
ModelType		ModelSize
		Correctness		Citation	Correctness		Citation	Correctness	Citation
		EMRec.	Rec.	Prec. F1.	Claim	Rec.	Prec. F1	Rec.-5 Prec.	Rec. Prec.	F1
Prompting-based
ChatGPT - 40.37 72.81 69.69 71.22 12.47 49.44 47.05 48.22 20.28 19.84 19.06 22.03 20.44
7B 24.32 17.24 17.87 17.55 4.53 3.92 5.38 4.54 12.56 11.32 6.03 6.35 6.19
LLaMA-2 13B 27.99 16.45 19.04 17.65 7.77 8.49 8.43 8.46 18.00 12.39 5.45 5.74 5.59
70B 31.53 44.18 44.79 44.48 10.43 23.75 22.43 23.07 18.50 14.79 10.10 10.50 10.30
7B 29.93 55.99 51.66 53.74 12.47 19.90 15.48 17.41 17.96 19.74 9.58 9.68 9.63
LLaMA-2-Chat 13B 34.39 37.15 38.17 37.65 13.83 16.50 16.09 16.29 21.34 18.86 8.94 9.06 9.00
70B 41.24 60.19 61.16 60.67 13.30 36.63 36.63 36.63 22.62 18.04 13.49 13.98 13.73
Vicuna-v1.5 7B 38.34 48.37 44.63 46.42 12.30 29.81 22.45 25.61 14.22 14.74 11.26 11.64 11.45
13B 35.20 51.92 53.40 52.65 14.33 31.15 28.99 30.03 22.06 19.60 13.04 13.74 13.38
7B 29.46 23.12 25.45 24.23 8.47 16.04 16.32 16.18 16.96 15.98 7.50 7.76 7.63
Mistral
8×7B 36.30 32.72 34.49 33.58 10.43 26.11 25.09 25.59 18.18 15.63 9.72 10.20 9.95
7B 38.57 64.90 59.67 62.18 11.07 49.25 42.69 45.74 17.52 21.29 17.56 18.53 18.03
Mistral-Instruct 8×7B 44.11 61.80 63.27 62.53 13.93 49.28 48.34 48.81 20.12 19.64 19.27 20.38 19.81
Post-hocRetrieval
ChatGPT - 37.68 27.11 27.05 27.08 18.77 14.55 14.55 14.55 25.14 22.85 12.29 12.29 12.29
LLaMA-2-Chat 70B 29.68 24.51 24.51 24.51 16.03 12.93 12.93 12.93 17.90 14.45 9.05 9.05 9.05
Mistral-Instruct 8×7B 33.90 24.57 24.48 24.52 17.37 15.68 15.68 15.68 24.16 18.28 9.78 9.78 9.78
Training-based
7B 29.96 67.82 66.97 67.39 6.90 22.34 32.40 26.45 2.34 1.98 10.53 18.80 13.50
Self-RAG(LLaMA-2)
13B 31.66 71.26 70.35 70.80 6.07 30.46 40.20 34.66 1.90 1.33 12.79 20.90 15.86
7B 40.32 67.67 63.67 65.61 9.63 42.30 40.06 41.15 12.86 21.09 21.35 21.36 21.35
VANILLA-SFT(LLaMA-2) 13B 40.85 71.49 66.21 68.75 10.27 46.75 44.47 45.58 12.68 22.80 23.64 23.71 23.67
FRONT(LLaMA-2) 7B 40.84 77.70 69.89 73.59 9.18 58.60 55.33 56.92 11.50 21.38 24.74 24.84 24.79
13B 41.51 78.44 73.66 75.97 9.32 60.31 59.21 59.75 11.94 22.61 24.86 25.39 25.12
Table1: MainresultsontheALCEbenchmark. Boldnumbersindicatethebestperformance,while_indicatesthe
second-bestperformance.
substringsofthegeneration. RegardingtheELI5 Mistral-7B(Jiangetal.,2023)toMistral-8x7B-
dataset,correctnessismeasuredthroughclaimre- MoE (Jiang et al., 2024). Regarding SFT LLMs,
call (Claim), evaluating whether the model’s re- weselecttheSFTcounterpartsoftheopen-source
sponse entails the ground truth sub-claims. For foundationalLLMsweused. Detailedprompting
theQAMPARIdataset,correctnessisassessedus- settingscanbefoundinAppendixC.
ingexactmatchprecision(Prec.)			andtop-5exact
-----------------------------	---	---	-------------	---	---	---	---	---	---	---
match recall (Rec.-5) — considered 100% if the 4.3.2 Post-hocRetrievalMethods.
predictionincludesatleastfivecorrectanswers.
					Following	Gao	et al.	(2023b),	we first	instruct
---	---	---	---	---	---------	---	------	--------	--------	--------
LLMstoanswerthegivenqueryinaclosed-book
4.3 Baselines
					setting,	and then	integrate	citations	in a post-hoc
---	---	---	---	---	--------	--------	---------	---------	-------------	---
Wecompareourmethodwiththreetypesofbase-
					manner.	Foreachgeneratedstatement,weemploy
------	----------------	--------	----------	---	-------	----------------------------------	---	---	---	---
lines:	prompting-based,	post-hoc	retrieval,	and
GTR(Nietal.,2022)toidentifyandcitethemost
training-based.
relevantdocumentfromthetop100retrieveddoc-
4.3.1 Prompting-basedMethods. uments. Weutilizethesamemodelsmentionedin
WedirectlypromptLLMsusingfew-shotdemon- prompting-basedsettingsforthisbaseline.
strations,eachconsistingofaquery,thetop5rel-
					4.3.3	Training-basedMethods.
---	---	---	---	---	-----	----------------------	---	---	---	---
evantretrieveddocuments,andananswerwithin-
linecitations.	Ourexperimentsencompassarange
--------------	-----------------------------	---	---	---	--------	---------------------------------	---	---	---	---
					Self-RAG	(Asaietal.,2023)Self-RAGtrainsthe
ofLLMs,fromfoundationalmodelstosupervised
					LLM	to learn	to adaptively	retrieve	passages	on-
---------------------	---	--------------------	---	---	---	--------	-------------	--------	--------	---
fine-tuning(SFT)LLMs.		ForfoundationalLLMs,
demandandenableittoreflectonitsgenerationto
we select	GPT-3.5-Turbo5	as the	representative
---------	--------------	------	--------------	---	---	---	---	---	---	---
furtherimprovegenerationqualityandattributions.
closed-sourcemodel,recognizedforitsnotableper-
formance.	Among	the open-source	foundational
---------	-----	---------------	------------	---	-----------	---	--------------------------	---	---	---
					VANILLA-SFT		Wedirectlyemploysupervised
LLMs,wefocusontheLLaMA-2seriesincluding
fine-tuningtotraintheLLMonourgeneratedtrain-
LLaMA2-7B,LLaMA2-13B,andLLaMA2-70B,
					ing data.	Given	a query	and corresponding		doc-
------------------------------------------------	--------------	-------------	-----	----	---------------------	-------	-------	-----------------	--------	--------
as well	as the Mistral	series, which	spans	from
					uments,	the LLM	is	required to	directly	generate
5Specifically,weutilizegpt-3.5-turbo-1106version					answerswithcitations.

Weimplement		FRONTwithdifferentsizesoffoun-					60				Pre-filtered
Post-filtered
dational	models	(LLaMA-2-7B		and	LLaMA-2-
--------	------	-----------	---	---	--------	---	---	---	---	---	---	---	---
)%(1FnoitatiC 50
13B)toevaluateitseffectiveness.				Duringtheeval-
-------------------------------	---	---	---	--------------	---	---	---	---	---	---	---	---	---
40
uation,FRONTutilizethesameretrievalsettingsas
those outlined		by Gao	et al.	(2023b).	Additional		30
--------------	-----------	------	----------	--------	----------	------	---	---	---	---	---	---	---
details	of training	and	evaluation	settings		can be	20
foundinAppendixD.
								ASQA		ELI5		QAMPARI
---	---	---	---	---	---	---	---	----	---	----	---	-------	---
5 ResultsandAnalysis
							Figure4:	AblationStudyonDataFiltering.
---	---	---	---	---	---	---	--------	-----------------------------	---	---	---	---	---
5.1 OverallResults
Simplysupervisedfine-tuningcanboostcitation
							FRONT	demonstrates		remarkable		generaliza-
-----------	------------------------------	---	--------------	---	----------	-----	----------	------------	---	----------	---	-----------	-------
quality.	AsshowninTable1,teachingLLMsto
							tion.	Compared	to	the varied		queries	and an-
generate	responses		with citations	via	supervised
							swer types	present	in	the ALCE		benchmark,	our
fine-tuning	significantly		enhances		citation	qual-
trainingdata,derivedexclusivelyfromtheAQuA-
ity,demonstratingsubstantialimprovementsover
MuSedataset(Kulkarnietal.,2020),exhibitsout-
bothprompt-basedandpost-hocretrievalbaselines
							of-domain	characteristics.			Nonetheless,		FRONT
------	-------------	---	-------------	----	--------	---	------------	----------------	---	--------	------------	---	---------
across	all datasets.		Specifically,	with	LLaMA-2-
							demonstrates	superior		citation	quality,		affirming
7B, VANILLA-SFTledtosubstantialgainsincita-
itsexceptionalabilitytogeneralizeacrossdiverse
tion F1	scores	over	prompting:	ASQA	(17.55	→
-------	------	----	----------	----	------	---	--------------------------------	---	---	---	---	-------------	---
							querytypesandretrievaldocuments.					Additionally,
65.61),ELI5(4.54→41.15),andQAMPARI(6.19
							while not	specifically		optimized		for correctness,
--------	-----------------------------------	---	---	---	---	---	---------	--------------	---	---------	------------	----------------	---
→21.35).	Thesegainshighlighttheeffectiveness
							FRONT	also showcases		modest	improvements		in
ofourtrainingdatagenerationpipeline.
thismetricoverVANILLA-SFTontheASQAand
							QAMPARIdatasets.			However,FRONTencounters
---	---	---	---	---	---	---	----------------	---	---	-----------------------	---	---	---
FRONT achievessignificantperformancegains
lowerRec.-5ontheQAMPARIdataset,likelydue
andsurpassesChatGPT.				While VANILLA-SFT
--------------------	---	---	---	-----------------	---	---	---	---	---	---	---	---	---
tothenatureofitsanswers,whichconsistofcon-
demonstrates		strong	performance,		it still	shows
------------	---	------	------------	---	--------	-----	---	---	---	---	---	---	---
catenatedentities,divergingsignificantlyfromour
notable	discrepancies		compared	to	leading	open-
-------	-------------	---	--------	---	-------	-----	---	---	---	---	---	---	---
trainingdata.
sourceLLMs,suchasMixtral-8×7B-Instruct(e.g.,
41.15 vs.	45.74)	and	ChatGPT	(e.g.,	41.15	vs.
---------	------	----	--------	-----------	-----	--------	-----------------	---	---	---	---	---	---
48.22)	on the	ELI5	dataset.	FRONT	not	only	5.2 AblationStudy
bridges	these	gaps	but also	establishes		signifi-
Weconductablationstudiestoverifytheeffective-
cant leads	across	all	datasets.	Specifically,		us-
---------------	------	---	---------	---------------	---	----	-----------------------------------	---	---	---	---	---	------
							nessofdifferentcomponentsproposedin						FRONT.
ing LLaMA-2-7B,			FRONT	comprehensively		out-
performsChatGPT,achievingincreasesof3.32%,
							EffectsofDataGenerationPipeline.						Asillus-
-------	---	------	-----------	-------	---	------	--------------------------------	--------	------	---	--------	------	--------
18.04%,	and	21.28%	in citation	quality		on the
							trated	in §5.1,	simply	SFT	achieves	strong	per-
ASQA,ELI5,andQAMPARIdatasetsrespectively.
							formance,	underscoring		the	high	quality	of our
---	---	---	---	---	---	---	---------	------------	---	---	----	-------	------
Thisperformanceunderscorestheeffectivenessof
							training	data.	Furthermore,		data	filtering,	a cru-
---	---	---	---	---	---	---	--------	-----	------------	---	----	----------	------
FRONTinenhancingattributioncapabilities.
							cial component		of our	data	generation		pipeline,
--------------------------------------	---	---	---	---	---	---	--------------	-------	-------	--------	----------	-------	---------
							plays a	pivotal	role in	ensuring	the	quality	of the
FRONTexhibitsscalabilitywithmodelsize.						As
illustratedatthebottomofTable1,theperformance generated data by filtering out queries that yield
of FRONT in terms of citation quality shows no- non-informativeanswersorfailtomeetattribution
							criteria.	To validate		the effectiveness		of	our data
---	---	---	---	---	---	---	---------	-----------	---	-----------------	---	---	--------
tableimprovementswhenscalingfrom7Bto13B.
filteringstrategies,weconductedexperimentscom-
Specifically,weobserveimprovementsof3.23%in
ASQA,4.97%inELI5,and1.33%inQAMPARI. paringmodelsfine-tunedonbothpre-filteredand
This upward trend underscores the scalability of post-filtereddata. Theresults,depictedinFigure
4,confirmthatmodelstrainedonfiltereddataex-
FRONT withincreasingmodelsize,demonstrating
thepotentialofFRONTinleveragingtheincreased hibitanotableimprovementincitationqualityover
capabilitiesoflargerLLMsforfurtherperformance thosetrainedonunfiltereddata,achievingsuperior
attributionperformancewithreduceddatavolume.
gains.

				ASQA				ELI5			QAMPARI
Model
			Correctness		Citation		Correctness		Citation	Correctness		Citation
---	---	---	-----------	---	--------	---	-----------	---	--------	-----------	---	--------	---
EMRec. Rec. Prec. F1. Claim Rec. Prec. F1 Rec.-5 Prec. Rec. Prec. F1
FRONT-7B 40.84 77.70 69.89 73.59 9.18 58.60 55.33 56.92 11.50 21.38 24.74 24.84 24.79
SELF-GUIDE(w/oConsistency) 38.99 70.69 64.48 67.44 10.04 47.63 44.80 46.17 12.18 20.03 22.50 22.58 22.54
VANILLA-SFT(w/oGround) 40.32 67.67 63.67 65.61 9.63 42.30 40.06 41.15 12.86 21.09 21.35 21.36 21.35
FRONT-13B 41.51 78.44 73.66 75.97 9.32 60.31 59.21 59.75 11.94 22.61 24.86 25.39 25.12
SELF-GUIDE(w/oConsistency) 40.99 73.08 68.13 70.52 10.06 50.68 49.78 50.23 13.94 22.38 23.73 23.99 23.85
VANILLA-SFT(w/oGround) 40.85 71.49 66.21 68.75 10.27 46.75 44.47 45.58 12.68 22.80 23.64 23.71 23.67
Table2: AblationstudyontheimpactofdifferenttrainingstageswithintheALCEbenchmark.
44.42
Citation-F1
				46.17					33		FRONT-13B
--------------	---	-----	-----	-----	-----------	---	---	-------------------	---------------	---	---------	------------	---
											F	-7 B
		43.71			SELF-GUIDE						R O	N T
Citation-Prec.									LLaMA2-70B-Chat		SEL F- G	U I D E -13B
			44.80		PROMPTGUIDE			ssenlufhtiaF 22..55
SELF-GUIDE-7B
45.16
Citation-Rec.											Mixtral-8x7B-Inst
-------------	---	-----	---	-----	-----	---	---	---	---	----------	-----------------	---	---
					47.63				22	Vicuna-13B
		44.00		46.00	48.00
LLaMA2-7B-Chat
Figure5:	Ablationstudyofdifferentgroundingguid-							11..55
--------	--------------------------------------	---	---	---	---	---	---	------	---	---	---	---	---
ancestrategiesontheELI5dataset.
									2200	3300	4400 5500	6600
---	---	---	---	---	---	---	---	---	----	----	---------	----	---
CitationF1
EffectsofGroundingGuidedGeneration(G3).
G3
empowers LLMs to first select relevant fine- Figure6: TherelationshipbetweencitationF1andhal-
grainedquotes,whichsubsequentlyguidethegen- lucination: Models positioned closer to the top-right
eration process. These quotes can provide fine- cornerexhibithighercitationqualityandalowerdegree
ofhallucination.
grainedsupervisionsignalsforattributedtextgen-
G3,
eration.	To evaluate		the effectiveness		of	we
---------------------------------------	-----------	---	-----------------	---------	----------	-----	-------	--------	-----------	---	---------	---	------
conduct	an ablation		study	comparing	it against
							eration	process.	Experiments		conducted		on the
twovariantswithdistincttrainingrecipes.						Given
ELI5datasetusingtheLLaMA-2-7Bmodelshow
that FRONTconsistsoftwostages,werefertothe that SELF-GUIDE outperforms PROMPT-GUIDE.
modeltrainedonlyduringthefirststage(without
ResultsdepictedinFigure5indicatethattraining
consistency-awarealignment)asSELF-GUIDE.						We
----------------------------------------	----------	---	-------	---	-----------	---	----------	---	-------------	---------	--------	--------------	------
							models	to	self-generate		grounded	quotes	before
firstcompare	SELF-GUIDE		against		VANILLA-SFT
							generating		attributed	responses	is	more effective
(w/oGround),whichistrainedtodirectlygenerate thansimplyincorporatingthesegroundedquotes
responseswithcitations,bypassingthegrounding
intotheprompt.
step. The	ablation	study,	detailed		in Table	2, re-
---------	--------	------	--------	---	--------	------	---	---	---	---	---	---	---
veals that models incorporating grounding guid- EffectsofConsistency-AwareAlignment(CCA).
ancesignificantlyoutperformtheir VANILLA-SFT The primary goal of CCA is to enhance the con-
counterpartsthatlacksuchgroundingmechanisms. sistency between grounded quotes and attributed
This highlights the crucial role of grounding in answers, thereby alleviating hallucinations and
							achievingmorepreciseattribution.					Toevaluatethis,
---	---	---	---	---	---	---	--------------------------------	---	---	---	---	---------------	---
bolsteringattribution.
Moreover, we explore an alternative variant we compare models that underwent only the G3
of grounding guidance. Considering that SELF- stage(SELF-GUIDE)withthosefurtherenhanced
GUIDE leverages the model itself to both select through the CCA stage (FRONT). As illustrated
grounded quotes and generate attributed answers inTable2,FRONTsignificantlyimprovescitation
in an end-to-end paradigm, a natural variant in- qualityoverSELF-GUIDE,demonstratingtheeffec-
volves breaking down this task into two distinct tivenessoftheCCAstageinenhancingattribution.
stages. Inthisvariant,ChatGPTistaskedwithex- Furthermore,toassessCCA’simpactonreduc-
tractinggroundedquotes. Subsequently,aseparate inghallucinations,weutilizeQAFactEval(Fabbri
modelistrainedtoutilizethesegroundedquotes, etal.,2022),awidelyusedmetricforfactualcon-
along with the query and retrieval documents, to sistency, which scores the consistency of model
directly output the response and citations. This responses to given documents on a scale from 0
variant,referredtoasPROMPT-GUIDED,integrates to5,withhigherscoresindicatinggreaterfaithful-
groundedquotesintotheprompttoguidethegen- ness. Specifically,weanalyzetheperformanceof

FRONTand	SELF-GUIDE		ontheELI5dataset.		As
						ChatGPT				0.93	4.08
showninFigure6, FRONTproducesmorefaithful FRONT-7BonASQA 0.94 3.86
outputsthanSELF-GUIDE,significantlyreducing FRONT-7BonELI5 0.92 3.96
hallucinations.						FRONT-7BonQAMPARI				0.86	3.62
---------------	---	---	---	---	---	-----------------	---	---	---	----	----
Table3: Humanevaluationonthequalityofgrounded
EffectsofTrainingDataScale.				Weanalyzethe
---------------------------	---	---	---	------------	---	---	---	---	---	---	---
quotes.
impact	of the	data scale	on model	performance
------	------------	----------	--------	-----------	-------	---	---	---	---	---	---
across	two training	stages.	In	particular,	we ran-
tors. Theresultsofthehumanevaluationindicate
domlysampled2k,4k,6k,and8kinstancesfrom
thatbothquotesextractedbyChatGPTandthose
our full	training	data	across two	distinct	training
--------	--------	----	----------	--------	--------	---	---	---	---	---	---
stages. These subsets were then utilized to fine- generated by FRONT are of high quality, further
tunevarious7Bmodelvariants,enablingacompar- substantiatingtheeffectivenessofourmethod.
ativeanalysisofperformancebasedondatascale.
Results	are shown	in	Figure	7, which	indicates
-------	---------	---	------	--------	---------	---	---	---	---	---	---
7 Conclusion
thatincreasingdatasizeshowssignificantenhance-
mentsincitationquality,indicatingapositivecor- Inthiswork,wepresent FRONT,atwo-stagetrain-
relationbetweendatasizeandmodelperformance. ingframeworkdesignedtoequipLLMswithfine-
AsFRONTimplementsanautomatedprocedureca-
						grained	attribution	capabilities.		FRONT	enables
---	---	---	---	---	---	-------	-----------	-------------	---	-----	-------
pableofgeneratinghigh-qualityattributeddataand LLMstoinitiallyselectsupportingquotes,which
constructingcontrastivesupervisionfromweakand thenguidethegenerationprocess. Byfurtheren-
strongLLMs,itholdsthepotentialforcontinuous
						hancing	the consistency		between		the grounding
---	---	---	---	---	---	-------	---------------	---	-------	---	-------------
performanceimprovements. andgenerationprocessviapreferenceoptimization,
thesesupportingquotescanserveasfine-grained
6 HumanEvaluation
						citations.	Through	comprehensive			experiments,
---	---	---	---	---	---	----------	-------	-------------	---	---	------------
Given the significant impact of the quality of FRONThasdemonstrateditsabilitytogeneratesu-
						perior grounded		responses		and highly	supportive
--------	------	---	------------	------------	---	---------------	------------------------	---------	---	----------	----------
grounded	quotes	on	fine-grained	verification	for
						citations.	Furtheranalysisshowsthat				FRONT sig-
users,weconductedahumanevaluationtoassess
thequalityofgroundedquotesatdifferentstagesof nificantlyreduceshallucinationsandbenefitsuser
verification.
ourframework:		(1)QuotesextractedbyChatGPT
----------------------------------	---------	---------------------------	-----------	---	------------	------------	---	---	---	---	---
from50sampleddatapointsduringtheG3					stage.
(2) Quotes	generated		by FRONT-7B		across three	8 Limitation
datasets,with50datapointssampledfromeach.
Weengagedfourannotators,eachwithrelevant Our study presents several limitations worth not-
expertiseandholdingatleastabachelor’sdegree. ing. Firstly, the validation of our framework is
Thequalityofquoteswasevaluatedontwodimen- predominantly conducted on models of sizes 7B
sions: authenticityandhelpfulness. Authenticity and13B,leavingtheexplorationoflargermodels,
(abinaryscaleof0/1)referstowhetherthequotes suchasLLaMA-270Bduetocomputationalcon-
genuinelyoriginatefromthecorrespondingdocu- straints. Secondly,ourframeworkreliesonaprior
ments(quotesthatarehallucinatedormismatched retrievalprocess,whereinrelevantdocumentsare
withthecorrespondingdocumentIDareconsidered retrieved at one time. The incorporation of adap-
inauthentic). Helpfulness (5-point Likert scale) tiveretrieval,enablingmoredynamicinteractions
referstothedegreetowhichthequotesarebenefi- withLLMs,couldpotentiallyenhanceperformance.
cialinaddressingthequery. TheresultsinTable3 Weleaveitforfutureresearch. Lastly,evaluating
representtheaveragescoresforallquoteswithin the correctness of long-form question answering
eachmodelresponse,withtwoannotatorsevaluat- presents inherent challenges, leading our frame-
ingeachresponsetoensurereliability. work to primarily enhance citation quality, with
Furthermore, to evaluate the consistency of modest advancements in correctness. Therefore,
quotequalityannotations,wecomputedtheinter- we advocate for the development of more robust
annotator agreement using Fleiss’ Kappa coeffi- metricscapableofaccuratelyassessingthecorrect-
cient. The obtained Kappa coefficient of 0.82 in- nessoflong-formQAresponses, pavingtheway
dicates	a high	level of	agreement	among	annota-	forfuturework.
-------	------	--------	---------	-----	-------	--------------	---	---	---	---	---

							rization.	InProceedingsofthe2022Conferenceof
We appreciate Yuchun Fan for providing valu- theNorthAmericanChapteroftheAssociationfor
							ComputationalLinguistics:			HumanLanguageTech-
-----------------	---	---------	---	---	-------	--------	-------------------------	---	---	------------------	---
able suggestions.		Xiaocheng			Feng is	the cor-
nologies,NAACL2022,Seattle,WA,UnitedStates,
responding	author	of	this	work.	We	thank the
----------	------	---	----	-----	---	---------	---	---	---	---	---
July10-15,2022,pages2587–2601.Associationfor
anonymous	reviewers		for	their	insightful	com-
---------	---------	---	---	-----	----------	----	---	---	---	---	---
ComputationalLinguistics.
ments.	This	work	was	supported	by	the Na-
------	----	----	---	---------	---	-------	---	---	---	---	---
AngelaFan,YacineJernite,EthanPerez,DavidGrang-
tional Key	R&D	Program		of China	via	grant No.
--------------	---	-------	--------	--------	---------	---------	----------	--------	-----------	--------------	-----
							ier, Jason	Weston,	and Michael	Auli. 2019.	ELI5:
2021ZD0112905,		the	National		Natural	Science
							long form	question	answering.	In Proceedings	of
Foundation	of	China	(NSFC)	(grant	62276078,
the57thConferenceoftheAssociationforCompu-
U22B2059), the Key R&D Program of Hei- tationalLinguistics,ACL2019,Florence,Italy,July
longjiang via grant 2022ZX01A32, the Interna- 28-August2,2019,Volume1: LongPapers,pages
3558–3567.AssociationforComputationalLinguis-
tionalCooperationProjectofPCL,PCL2022D01
tics.
andtheFundamentalResearchFundsfortheCen-
tralUniversities(GrantNo.HIT.OCEF.2023018). ZhangyinFeng,WeitaoMa,WeijiangYu,LeiHuang,
HaotianWang,QianglongChen,WeihuaPeng,Xi-
							aochengFeng,BingQin,andTingLiu.2023.				Trends
---	---	---	---	---	---	---	------------------------------------	---	---	---	------
inintegrationofknowledgeandlargelanguagemod-
References
els:Asurveyandtaxonomyofmethods,benchmarks,
CoRR,abs/2311.05876.
Samuel	Joseph	Amouyal,	Ohad	Rubin,		Ori Yoran,	andapplications.
--------	---------	--------	---------	-------	--------------	----------	----------------	------	-----------------	--------	---------
Tomer	Wolfson,	Jonathan		Herzig,	and	Jonathan
							Luyu Gao,	Zhuyun	Dai, Panupong	Pasupat,	Anthony
Berant.	2022.	QAMPARI:		:	An open-domain
							Chen,	Arun	Tejasvi Chaganty,	Yicheng	Fan, Vin-
question	answering		benchmark		for questions	with
centY.Zhao,NiLao,HongraeLee,Da-ChengJuan,
many	answers	from	multiple	paragraphs.		CoRR,
----	-------	----	--------	-----------	---	-----	----------	----	------	-----------------	---
							and Kelvin	Guu.	2023a.	RARR: researching	and
abs/2205.12665.
revisingwhatlanguagemodelssay,usinglanguage
AkariAsai,ZeqiuWu,YizhongWang,AvirupSil,and models. InProceedingsofthe61stAnnualMeeting
oftheAssociationforComputationalLinguistics(Vol-
Hannaneh	Hajishirzi.		2023.	Self-rag:	Learning	to
--------	-----------	---	-----	---------	--------	---	-----	-----------------------------------	---	---	---
							ume1:	LongPapers),ACL2023,Toronto,Canada,
retrieve,generate,andcritiquethroughself-reflection.
July9-14,2023,pages16477–16508.Associationfor
CoRR,abs/2310.11511.
ComputationalLinguistics.
Yuntao Bai,	Andy	Jones,	Kamal	Ndousse,		Amanda
-----------	----	------	-----	--------	---	------	---	---	---	---	---
TianyuGao,HowardYen,JiatongYu,andDanqiChen.
Askell,	AnnaChen,		NovaDasSarma,		DawnDrain,
---------	---------	----	-------------	---	----------	---------	------------------	-------------------------------------	--------------------	---	---
							2023b.	Enablinglargelanguagemodelstogenerate
Stanislav	Fort,	Deep	Ganguli,		Tom	Henighan,
							textwithcitations.		CoRR,abs/2305.14627.
NicholasJoseph,SauravKadavath,JacksonKernion,
TomConerly,SheerElShowk,NelsonElhage,Zac
							Yunfan Gao,	Yun	Xiong, Xinyu	Gao, Kangxiang	Jia,
---------------	---	-----	----------	---	-------	-----	-----------	---	------------	--------------	----
Hatfield-Dodds,		Danny	Hernandez,		Tristan	Hume,
JinliuPan,YuxiBi,YiDai,JiaweiSun,QianyuGuo,
ScottJohnston,ShaunaKravec,LianeLovitt,Neel Meng Wang, and Haofen Wang. 2023c. Retrieval-
Nanda,	Catherine	Olsson,		Dario	Amodei,	Tom B.
--------	---------	-------	---------------	---------	-------	----------	------------------------------------------	--------------------	---	---	---
							augmentedgenerationforlargelanguagemodels:				A
Brown,	Jack	Clark,	Sam McCandlish,		Chris	Olah,
							survey.	CoRR,abs/2312.10997.
Benjamin	Mann,	and	Jared	Kaplan.	2022.	Train-
ing a	helpful	and	harmless	assistant		with rein-
SylvainGugger,LysandreDebut,ThomasWolf,Philipp
forcement learning from human feedback. CoRR, Schmid,ZacharyMueller,SourabMangrulkar,Marc
abs/2204.05862. Sun,andBenjaminBossan.2022. Accelerate: Train-
ingandinferenceatscalemadesimple,efficientand
BerndBohnet,VinhQ.Tran,PatVerga,RoeeAharoni,
							adaptable.		https://github.com/huggingface/
------	------	-----	-------	-------	-----	------	----------	---	-------------------------------	---	---
Daniel	Andor,	Livio	Baldini	Soares,	Jacob	Eisen-
accelerate.
stein,KuzmanGanchev,JonathanHerzig,KaiHui,
TomKwiatkowski,JiMa,JianmoNi,TalSchuster, OrHonovich, RoeeAharoni, JonathanHerzig, Hagai
WilliamW.Cohen,MichaelCollins,DipanjanDas, Taitelbaum,DoronKukliansy,VeredCohen,Thomas
Donald Metzler, Slav Petrov, and Kellie Webster. Scialom, Idan Szpektor, Avinatan Hassidim, and
2022.	Attributed	question		answering:		Evaluation
-----	----------	--------	---	----------	---	----------	-----	-------	-----------	-------------	-------
							Yossi	Matias.	2022. TRUE:	re-evaluating	factual
andmodelingforattributedlargelanguagemodels.
							consistencyevaluation.		InProceedingsofthe2022
---	---	---	---	---	---	---	----------------------	---	----------------------	---	---
CoRR,abs/2212.08037. Conference of the North American Chapter of the
							AssociationforComputationalLinguistics:				Human
---	---	---	---	---	---	---	---------------------------------------	---	---	---	-----
CanyuChenandKaiShu.2023. Combatingmisinfor- LanguageTechnologies,NAACL2022,Seattle,WA,
mation in the age of llms: Opportunities and chal- UnitedStates, July10-15, 2022, pages3905–3920.
lenges.	CoRR,abs/2311.05656.
-------	--------------------	---	---	---	---	---	---	---	---	---	---
AssociationforComputationalLinguistics.
Alexander R. Fabbri, Chien-Sheng Wu, Wenhao Liu, LeiHuang,WeijiangYu,WeitaoMa,WeihongZhong,
and Caiming Xiong. 2022. Qafacteval: Improved Zhangyin Feng, Haotian Wang, Qianglong Chen,

guagemodels:	Principles,taxonomy,challenges,and					Principles.
openquestions.	CoRR,abs/2311.05232.
						Patrick	S. H. Lewis,	Ethan	Perez, Aleksandra		Pik-
						tus, Fabio	Petroni,	Vladimir	Karpukhin,		Naman
Ziwei Ji, Nayeon	Lee,	Rita	Frieske,	Tiezheng	Yu,
DanSu,YanXu,EtsukoIshii,YejinBang,Andrea Goyal,HeinrichKüttler,MikeLewis,Wen-tauYih,
Madotto,andPascaleFung.2023. Surveyofhalluci- Tim Rocktäschel, Sebastian Riedel, and Douwe
nationinnaturallanguagegeneration. ACMComput. Kiela. 2020. Retrieval-augmented generation for
Surv.,55(12):248:1–248:38. knowledge-intensiveNLPtasks. InAdvancesinNeu-
						ralInformationProcessingSystems33:				AnnualCon-
---	---	---	---	---	---	----------------------------------	---	---	---	----------	---
AlbertQ.Jiang,AlexandreSablayrolles,ArthurMen-
ferenceonNeuralInformationProcessingSystems
sch,ChrisBamford,DevendraSinghChaplot,Diego 2020,NeurIPS2020,December6-12,2020,virtual.
de Las Casas,	Florian	Bressand,		Gianna	Lengyel,
-------------	-------	---------	---	------	--------	---	---	---	---	---	---
Guillaume Lample, Lucile Saulnier, Lélio Re- Dongfang Li, Zetian Sun, Xinshuo Hu, Zhenyu Liu,
nard Lavaud, Marie-Anne Lachaux, Pierre Stock, Ziyang Chen, Baotian Hu, Aiguo Wu, and Min
TevenLeScao,ThibautLavril,ThomasWang,Timo- Zhang. 2023. A survey of large language models
						attribution.	CoRR,abs/2311.03731.
-----------------------------------	---	---	---	---	-------	------------	--------------------	---	---	---	---
théeLacroix,andWilliamElSayed.2023.					Mistral
7b. CoRR,abs/2310.06825.
						Nelson F.	Liu, Tianyi	Zhang,	and Percy	Liang.	2023.
---	---	---	---	---	---	---------	-----------	------	---------	------	-----
Albert Q. Jiang, Alexandre Sablayrolles, Antoine Evaluatingverifiabilityingenerativesearchengines.
Roux,ArthurMensch,BlancheSavary,ChrisBam- InFindingsoftheAssociationforComputationalLin-
ford,DevendraSinghChaplot,DiegodeLasCasas, guistics: EMNLP2023,Singapore,December6-10,
Emma Bou Hanna, Florian Bressand, Gianna 2023,pages7001–7025.AssociationforComputa-
Lengyel,	Guillaume	Bour,	Guillaume		Lample,	tionalLinguistics.
------------	---------	------	---------	---------	-------	------------------	---	---	---	---	---
Lélio Renard	Lavaud,	Lucile		Saulnier,	Marie-
AnneLachaux,PierreStock,SandeepSubramanian, JacobMenick,MajaTrebacz,VladimirMikulik,John
Sophia Yang, Szymon Antoniak, Teven Le Scao, Aslanides, H. Francis Song, Martin J. Chadwick,
Théophile Gervet, Thibaut Lavril, Thomas Wang, Mia Glaese, Susannah Young, Lucy Campbell-
						Gillingham,	Geoffrey		Irving, and Nat	McAleese.
---------------------------------------	---	---	---	---	----	-----------	--------	---	---------------	---------	---
TimothéeLacroix,andWilliamElSayed.2024.					Mix-
2022. Teachinglanguagemodelstosupportanswers
tralofexperts.	CoRR,abs/2401.04088.
---------------	--------------------	-------	-----	------	------	-------------------	---	--------------------	---	---	---
						withverifiedquotes.		CoRR,abs/2203.11147.
Ehsan Kamalloo,	Aref	Jafari,	Xinyu	Zhang,	Nandan
Thakur,andJimmyLin.2023. HAGRID:Ahuman- ReiichiroNakano,JacobHilton,SuchirBalaji,JeffWu,
llmcollaborativedatasetforgenerativeinformation- Long Ouyang, Christina Kim, Christopher Hesse,
seekingwithattribution. CoRR,abs/2307.16883. ShantanuJain,VineetKosaraju,WilliamSaunders,
						Xu Jiang,	Karl	Cobbe,	Tyna Eloundou,	Gretchen
---	---	---	---	---	---	---------	----	------	--------------	--------	---
Krueger,KevinButton,MatthewKnight,Benjamin
VladimirKarpukhin,BarlasOguz,SewonMin,Patrick
S.H.Lewis,LedellWu,SergeyEdunov,DanqiChen, Chess,andJohnSchulman.2021. Webgpt: Browser-
andWen-tauYih.2020. Densepassageretrievalfor assisted question-answering with human feedback.
open-domainquestionanswering. InProceedingsof CoRR,abs/2112.09332.
the2020ConferenceonEmpiricalMethodsinNat-
						Jianmo	Ni, Chen	Qu, Jing	Lu, Zhuyun	Dai,	Gus-
-------------	-----------	---	-----	-----	-------	------	--------	--------	----------	----	----
ural Language	Processing,		EMNLP	2020,	Online,
November16-20,2020,pages6769–6781.Associa- tavo Hernández Ábrego, Ji Ma, Vincent Y. Zhao,
tionforComputationalLinguistics. YiLuan,KeithB.Hall,Ming-WeiChang,andYinfei
						Yang.2022.	Largedualencodersaregeneralizable
---	---	---	---	---	---	----------	---------------------------------	---	---	---	---
SayaliKulkarni,SheideChammas,WanZhu,FeiSha, retrievers. InProceedingsofthe2022Conferenceon
and Eugene Ie. 2020. Aquamuse: Automatically EmpiricalMethodsinNaturalLanguageProcessing,
generatingdatasetsforquery-basedmulti-document EMNLP2022,AbuDhabi,UnitedArabEmirates,De-
summarization. CoRR,abs/2010.12694. cember7-11,2022,pages9844–9855.Association
forComputationalLinguistics.
TomKwiatkowski,	JennimariaPalomaki,				OliviaRed-
---------------	-------------------	---	---	---	----------	---	---	---	---	---	---
field,MichaelCollins,AnkurP.Parikh,ChrisAlberti, OpenAI. 2023. GPT-4 technical report. CoRR,
DanielleEpstein,IlliaPolosukhin,JacobDevlin,Ken-						abs/2303.08774.
------------------------------------------------	---	---	---	---	---	---------------	---	---	---	---	---
tonLee,KristinaToutanova,LlionJones,Matthew
LongOuyang,JeffreyWu,XuJiang,DiogoAlmeida,
Kelcey, Ming-Wei		Chang,	Andrew	M.	Dai, Jakob
----------------	--------	--------	-------	-----	----------	-------	--------------	---	---------------	---	-----
						Carroll	L. Wainwright,		Pamela Mishkin,		Chong
Uszkoreit,	Quoc Le,	and Slav	Petrov.	2019.	Natu-
ralquestions: abenchmarkforquestionanswering Zhang,SandhiniAgarwal,KatarinaSlama,AlexRay,
research. Trans.Assoc.Comput.Linguistics,7:452– JohnSchulman,JacobHilton,FraserKelton,Luke
466.						Miller,MaddieSimens,AmandaAskell,PeterWelin-
----	---	---	---	---	---	--------------------------------------------	---	---	---	---	---
der,PaulF.Christiano,JanLeike,andRyanLowe.
Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying 2022. Training languagemodelsto followinstruc-
Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. tionswithhumanfeedback. InAdvancesinNeural
Gonzalez, Hao Zhang, and Ion Stoica. 2023. Effi- InformationProcessingSystems35: AnnualConfer-
cientmemorymanagementforlargelanguagemodel enceonNeuralInformationProcessingSystems2022,

-December9,2022.							raker,BenZevenbergen,VinodkumarPrabhakaran,
							Mark Diaz,	Ben	Hutchinson,		Kristen	Olson, Ale-
AleksandraPiktus,FabioPetroni,VladimirKarpukhin, jandraMolina,ErinHoffman-John,JoshLee,Lora
DmytroOkhonko,SamuelBroscheit,GautierIzacard,
							Aroyo,	Ravi Rajakumar,		Alena	Butryna,	Matthew
---	---	---	---	---	---	---	------	---------------	---	-----	--------	-------
Patrick S. H. Lewis, Barlas Oguz, Edouard Grave, Lamm,ViktoriyaKuzmina,JoeFenton,AaronCo-
Wen-tauYih,andSebastianRiedel.2021. Theweb hen,RachelBernstein,RayKurzweil,BlaiseAgüera
isyouroyster-knowledge-intensiveNLPagainsta yArcas,ClaireCui,MarianCroak,EdH.Chi,and
verylargewebcorpus. CoRR,abs/2112.09924. QuocLe.2022. Lamda: Languagemodelsfordialog
CoRR,abs/2201.08239.
applications.
KrishnaPillutla,SwabhaSwayamdipta,RowanZellers,
JohnThickstun,SeanWelleck,YejinChoi,andZaïd
							Hugo Touvron,	Louis	Martin,	Kevin	Stone,	Peter Al-
---	---	---	---	---	---	---	-------------	-----	-------	-----	------	---------
Harchaoui.2021. MAUVE:measuringthegapbe- bert, Amjad Almahairi, Yasmine Babaei, Nikolay
tweenneuraltextandhumantextusingdivergence Bashlykov,SoumyaBatra,PrajjwalBhargava,Shruti
frontiers.	In	Advances	in Neural	Information		Pro-
----------	---	--------	---------	-----------	---	----	---	---	---	---	---	---
Bhosale,DanBikel,LukasBlecher,CristianCanton-
cessingSystems34:		AnnualConferenceonNeural
-----------------	---	------------------------	---	---	---	---	---	---	---	---	---	---
Ferrer,MoyaChen,GuillemCucurull,DavidEsiobu,
InformationProcessingSystems2021,NeurIPS2021,
JudeFernandes,JeremyFu,WenyinFu,BrianFuller,
December6-14,2021,virtual,pages4816–4828.
CynthiaGao,VedanujGoswami,NamanGoyal,An-
thonyHartshorn,SagharHosseini,RuiHou,Hakan
RonakPradeep,SahelSharifymoghaddam,andJimmy Inan,MarcinKardas,ViktorKerkez,MadianKhabsa,
Lin.2023.	Rankvicuna:		Zero-shotlistwisedocument
---------	-----------	---	-------------------------	---	---	---	---	---	---	---	---	---
IsabelKloumann,ArtemKorenev,PunitSinghKoura,
rerankingwithopen-sourcelargelanguagemodels.
Marie-AnneLachaux,ThibautLavril,JenyaLee,Di-
CoRR,abs/2309.15088.
anaLiskovich,YinghaiLu,YuningMao,XavierMar-
tinet,TodorMihaylov,PushkarMishra,IgorMoly-
Rafael Rafailov,		Archit	Sharma,	Eric	Mitchell,	Ste-
----------------	---	------	-------	----	---------	----	----------	----	------	--------	------	-------
							bog, Yixin	Nie,	Andrew	Poulton,	Jeremy	Reizen-
fanoErmon,ChristopherD.Manning,andChelsea
stein,RashiRungta,KalyanSaladi,AlanSchelten,
Finn.	2023.	Direct preference		optimization:		Your
------------------------------------	-----	-----------------	---	-------------	---	-----	--------------	-----	-------	------	------	---------
							Ruan Silva,	Eric	Michael	Smith,	Ranjan	Subrama-
languagemodelissecretlyarewardmodel.						CoRR,
							nian, Xiaoqing	Ellen	Tan,	Binh	Tang,	Ross Tay-
abs/2305.18290.
							lor, Adina	Williams,	Jian	Xiang	Kuan,	Puxin Xu,
---	---	---	---	---	---	---	----------	---------	----	-----	-----	---------
ZhengYan,IliyanZarov,YuchenZhang,AngelaFan,
Jeff Rasley,	Samyam	Rajbhandari,		Olatunji		Ruwase,
------------	------	------------	---	--------	---	-------	---	---	---	---	---	---
and Yuxiong He. 2020. Deepspeed: System opti- Melanie Kambadur, Sharan Narang, Aurélien Ro-
driguez,RobertStojnic,SergeyEdunov,andThomas
mizationsenabletrainingdeeplearningmodelswith
							Scialom.2023.	Llama2:		Openfoundationandfine-
-------------------------	---	---	---	---------	-------	---	----------------	-------	--------------------	----------------------	---	---
over100billionparameters.				InKDD’20:	The26th
							tunedchatmodels.		CoRR,abs/2307.09288.
ACMSIGKDDConferenceonKnowledgeDiscovery
and Data	Mining,	Virtual	Event,	CA,	USA,	August
--------	-------	-------	------	---	----	------	---	---	---	---	---	---
23-27,2020,pages3505–3506.ACM. TheodoraWorledge,JudyHanwenShen,NicoleMeis-
							ter,CalebWinston,andCarlosGuestrin.2023.					Uni-
---	---	---	---	---	---	---	----------------------------------------	---	---	---	---	----
fyingcorroborativeandcontributiveattributionsin
Freda Shi,	Xinyun	Chen,	Kanishka	Misra,		Nathan
----------	------	-----	--------	------	---	------	--------------------	---	---	--------------------	---	---
							largelanguagemodels.			CoRR,abs/2311.12233.
Scales,DavidDohan,EdH.Chi,NathanaelSchärli,
andDennyZhou.2023.			Largelanguagemodelscan
------------------	---	---	----------------------	---	---	---	---	---	---	---	---	---
beeasilydistractedbyirrelevantcontext. InInterna- Fangyuan Xu, Weijia Shi, and Eunsol Choi. 2023a.
tionalConferenceonMachineLearning,ICML2023, RECOMP:improvingretrieval-augmentedlmswith
23-29 July 2023, Honolulu, Hawaii, USA, volume compression and selective augmentation. CoRR,
abs/2310.04408.
202ofProceedingsofMachineLearningResearch,
pages31210–31227.PMLR.
ShichengXu,LiangPang,HuaweiShen,XueqiCheng,
IvanStelmakh,YiLuan,BhuwanDhingra,andMing- andTat-SengChua.2023b. Search-in-the-chain: To-
Wei Chang. 2022. ASQA: factoid questions meet wards the accurate, credible and traceable content
long-formanswers. InProceedingsofthe2022Con- generation for complex knowledge-intensive tasks.
CoRR,abs/2304.14732.
ferenceonEmpiricalMethodsinNaturalLanguage
Processing,EMNLP2022,AbuDhabi,UnitedArab
Emirates,December7-11,2022,pages8273–8288. Ori Yoran, Tomer Wolfson, Ori Ram, and Jonathan
AssociationforComputationalLinguistics. Berant. 2023. Making retrieval-augmented lan-
							guage models	robust	to	irrelevant	context.	CoRR,
-----	----------	------	-------------	--------	--------	-----	---------------	------	---	----------	--------	-----
Romal	Thoppilan,	Daniel	De	Freitas,	Jamie	Hall,	abs/2310.01558.
Noam	Shazeer,	Apoorv	Kulshreshtha,		Heng-Tze
Cheng,AliciaJin,TaylorBos,LeslieBaker,YuDu, Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang,
YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Xiaolei Wang, Yupeng Hou, Yingqian Min, Be-
AminGhafouri,MarceloMenegali,YanpingHuang, ichenZhang,JunjieZhang,ZicanDong,YifanDu,
MaximKrikun,DmitryLepikhin,JamesQin,Dehao Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao
Chen,YuanzhongXu,ZhifengChen,AdamRoberts, Jiang,RuiyangRen,YifanLi,XinyuTang,Zikang
MaartenBosma,YanqiZhou,Chung-ChingChang, Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen.
IgorKrivokon,WillRusch,MarcPickett,KathleenS. 2023. A survey of large language models. CoRR,
Meier-Hellstern,		Meredith	Ringel	Morris,		Tulsee	abs/2303.18223.
----------------	---	--------	------	-------	---	------	---------------	---	---	---	---	---

Chu, JingchangChen,		andBingQin.2024.		Anin-
lengthofanswersandtheaveragenumberofcita-
formationbottleneckperspectiveforeffectivenoise
tionsgeneratedforvarioustypesofquerieswithin
filteringonretrieval-augmentedgeneration.
ourdataset,asshowninTable4.
YutaoZhu,HuayingYuan,ShutingWang,JiongnanLiu,
Wenhan	Liu, Chenlong	Deng,	Zhicheng	Dou, and
------	-------------	-----	--------	--------	---	---	---	---	---	---
A.2 DetailsofDataFiltering
Ji-RongWen.2023.		Largelanguagemodelsforinfor-
----------------	---	----------------------------	--------------------	---	---	---	---	---	---	---
mationretrieval:		Asurvey.	CoRR,abs/2308.07107.
WetrainedourAttributedDiscriminatorusingthe
					manually	annotated	data	provided	by	Liu et al.
---	---	---	---	---	--------	---------	----	--------	---	----------
A DetailsofDataGenerationPipeline
					(2023),	which is	sampled	from	real generative
---	---	---	---	---	-------	--------	-------	----	---------------	---
A.1 DataStatistic searchengines. Eachstatementanditsciteddocu-
menthavebeenmeticulouslyannotatedforattribu-
#Questions 8,098 tion,categorizedintothreetypes: completesupport,
					partialsupport,andnosupport.				Fortraining,weuti-
------------	---	---	---	----	----------------------------	---	---	---	------------------	---
➥#LongAnswer				5667
lizedadatasetof8,834instances,comprising6,415
➥#ShortAnswer				2431
-------------	---	---	---	----	---	---	---	---	---	---
instancesofcompletesupport,1,552ofpartialsup-
Avg. WordsperAnswer 50.48 port, and 867 of no support. The discriminator
initializedwithLLaMA-2-7B,wastrainedwitha
➥Avg.	WordsperLongAnswer			69.15
-----	-------------------	---	---	-----	---------------------------	---	---	---	--------------	---
					maximumsequencelengthof512.				Wetraineditfor
➥Avg.	WordsperShortAnswer			6.94
3epochs,withatotalbatchsizeof128,andapeak
Avg. CitationperAnswer 4.40 learning rate of 2e-5, incorporating 3% warmup
steps,followedbyalineardecay.
➥Avg.	CitationperLongAnswer			4.68
-----	----------------------	---	---	----	------	-----------------	---------	---------	----------	-----
➥Avg.					During	the data	filtering	stage,	we first	break
	CitationperShortAnswer			3.77
					down	the automatically		generated	attributed	an-
swersintostatementformandusethetraineddis-
Table 4:	The statistics	of the	data generated	by our
--------	--------------	------	--------------	------	---	---	---	---	---	---
criminatortoannotatetheattributionbetweeneach
automaticdatagenerationpipeline.
					statement	and its	cited	documents.	Specifically,
---	---	---	---	---	---------	-------	-----	----------	-------------	---
weassigndifferentattributionscorestoeachstate-
Table4presentsthestatisticsofthedataautomat-
					ment s	based on	its attribution		relationship	with
-------------------------------------------	---	---	---	---	-----------------------------------	--------	---------------	---	------------	----
icallygeneratedbyourdatagenerationpipeline.				In
					citeddocumentsd,asshowninEquation7.					Con-
total,wecollected8,098questionsfromtheNatural
sequently,foreachattributedanswer,wecancal-
Questions(NQ)dataset,ofwhich5,667questions
					culateitsaverageattributionscore.				Attributedan-
---	---	---	---	---	---------------------------------	---	---	---	-------------	---
weregatheredfromthosewithlong-formanswers,
swerswithanaverageattributionscorebelow0.8
and2,431questionswerecollectedfromthosewith
					are filtered	out.	The threshold		of 0.8 was	deter-
---	---	---	---	---	------------	----	-------------	---	----------	------
short-formfactoidanswers.
minedthroughpreliminarytestingonthedevelop-
Forquestionsrequiringlong-formanswers,we
					ment set,	for which	we	manually	annotated	100
---	---	---	---	---	---------	---------	---	--------	---------	---
initializedourquerysourcewiththeAQUAMUSE
samplestoensuretheeffectivenessofourfiltering
dataset (Kulkarni		et al., 2020),	which consists	of
-----------------	---	--------------	--------------	---	---	---	---	---	---	---
criteria.
high-qualityqueriesspecificallydesignedforlong-
formresponseswithintheNQdataset,recognized
as“good”bythemajorityofNQevaluators.				Inthis
------------------------------------	---	---	---	------	---	---	---	---	---	---

						1, Dis(s,d)		= complete	support
--------------------------------------------	---	---	---	---	---	-----------	---	----------	-------	---
way,utilizingarefinedandsuperiorqualityquery						 
setlaidarobustgroundworkforourtrainingdata r(s)= 0.5,Dis(s,d) = partial support

generation,streamliningthedatafilteringprocess. 0, Dis(s,d) = no support
Forfactoidqueriesthatnecessitateshort-forman-										(7)
---------------------------------------------	---	---	---	---	---	---	---	---	---	---
swers,wedirectlysampledfromtheoriginalNQ
dataset, leveraging		its abundance	and	inherently
-------------------	---	-------------	---	----------	---	---	---	---	---	---
B DetailsofEvaluationMetrics
highquality.
During the data generation process, our initial In addition to evaluating citation quality and cor-
querysetcomprised7,725queriesrequiringlong- rectness,theALCEbenchmarkincludesabroader
formanswersand4,000queriesnecessitatingshort- setofdimensions,suchasfluency,ROUGE-L,and
form answers.		After a two-stage	data	filtering	generationlength.
-------------	---	-----------------	----	---------	-----------------	---	---	---	---	---

tably,wecalculatefluencyonlyfortheASQAand							Table9.
ELI5 datasets,		omitting	it for	QAMPARI,		as the
response in QAMPARI typically consists of lists D.2 RetrievalSettings
ofshortanswers.		ArelativelyhighMAUVEscore
---------------	---	-------------------------	---	---	---	---	---	---	---	---	---	---	---
Duringtheevaluation,weadoptthesameretrieval
indicatesthatthegenerationissufficientlyfluent.
							settings	as specified		by Gao	et	al. (2023b).	For
-------	---------------------------------	---	---	--------	---	------	--------------------------------	------------	-------	------	---------	------------	-------
							the ASQA	and	QAMPARI		datasets,	we	use the
ROUGE-L	Inadditiontoevaluatingthecorrect-
							denseretrieverGTR(Nietal.,2022).					FortheELI5
ness of	the model-generated			content,	we	employ
dataset,weemploythesparseretrieverBM25.
ROUGE-Ltoassesstheoverallqualityandtextual
coherenceoftheresponses.
E MoredetailaboutAblationStudy
C Prompts
E.1 TheEffectofTrainingDataScale.
C.1 PromptsforPrompting-basedMethods Weexaminehowmodelperformancevarieswith
							changesindatascale,asdepictedinFigure7.						The
---	---	---	---	---	---	---	---------------------------------------	---	---	---	---	---	---
FollowingGaoetal.(2023b),weadoptthevanilla
							upper part	of	the figure	illustrates		the	impact of
---	---	---	---	---	---	---	----------	---	----------	-----------	---	---	---------
promptingstrategyforitssimplicityandeffective-
							the training	data	scale	on	citation	quality	during
------------------------------------	------	--------	--------	--------	---------	---------	-------------	----	-------	----------	----------	--------	---------
ness. Specifically,		the	prompts	vary	according	to
							the Grounding		Guided	Generation		training	stage,
thetypeofdatawithintheALCEbenchmark.						For
							with datasets		ASQA,	ELI5,	and	QAMPARI	rep-
long-form	QA	datasets	such	as ASQA		and ELI5,
							resented	from	left to	right.	Similarly,		the lower
the prompt	format	is	detailed	in Table	5.	For the
partofthefiguredescribestheinfluenceduringthe
short-form	QA	dataset	QAMPARI,		the	format is
----------	---	-------	--------	---	---	---------	---	---	---	---	---	---	---
Consistency-AwareAlignmenttrainingstage.
outlinedinTable6.
E.2 TheGeneralizationAcrossModel
C.2 Instructionsfor			FRONT
-------------------	---	---	-----	---	---	---	---	---	---	---	---	---	---
Architectures.
Duringthetrainingprocess,wefollowtheinstruc-
							FRONT	demonstrates		exceptional		generalization
---	---	---	---	---	---	---	-----	------------	---	-----------	---	--------------	---
Alpaca6.
tion format	of		Specifically,		we	employ
-----------	---	---	-------------	---	---	------	---	---	---	---	---	---	---
capabilitiesacrossvariousfoundationalmodelar-
variedinstructionsfordifferentquestiontypes,as
							chitectures.	Specifically,transitioningthefounda-
---	---	---	---	---	---	---	------------	------------------------------------	---	---	---	---	---
delineatedinTable7forlong-formquestionsand
							tional model		from LLaMA-2-7B			to the	stronger
---	---	---	---	---	---	---	------------	---	---------------	---	---	------	--------
Table8forshort-formquestions.
							foundational		model,	Mistral-7B,		results	in even
---	---	---	---	---	---	---	------------	-----------	------	------------	---	-------	--------
							greater	performance		enhancements		as	shown in
D ExperimentalDetails
							Figure8.	Thisfurtherunderscoresthebroadappli-
---------------------	---	---	-----	---	---	---	-----------------------------	------------------------------------	---	---	------	---	---
							cabilityandgeneralizabilityof				FRONT.
D.1 TrainingDetailsof			FRONT
Thetrainingofallmodelsisexecutedon4Nvidia E.3 Theeffectofβ inConsistency-Aware
A100GPUs,eachwith80GBofmemory,leverag- AlignmentTrainingStage
ingtheDeepspeed(Rasleyetal.,2020)andHug-
							In the	Consistency-Aware			Alignment		Training
---	---	---	---	---	---	---	------	-----------------	---	---	---------	---	--------
gingFaceAcceleratelibraries(Guggeretal.,2022)
							Stage,	the β	parameter	in	Direct	Preference	Op-
--------------------------------------	---	---	---	---	---	-----	----------	-----	---------	---	------	----------	------
toconductmulti-GPUdistributedtraining.						Given
							timization	(DPO)	controls		the	strength	of the
thelongnatureoftheinputs,themaximumtoken
Kullback-Leiblerpenalty,typicallysetwithinthe
lengthissetto2,048tokens.
							range of	0.1 to	0.5.	A higher	β	value	indicates a
------	---	---------	-----	----------	---	------	--------	------	----	--------	---	-----	-----------
During	the	grounding	guide	generation		stage,
preferenceforthepolicymodel’strainingprocess
modelsaretrainedfor5epochswithatotalbatch
toremainclosertotheinitiallyreferencedmodel.
size of	128, a peak	learning		rate of	2e-5	with 3%
--------------------------------------	-----------	--------	---	-------	----	-------	-----------------------------------	------	---	---	-----	------	--------
							In extreme	cases,	as	β →	0, we	ignore	the con-
warmupstepsfollowedbyalineardecay.						During
							straintsimposedbythereferencemodel.						Thisset-
thecontrastivealignmentstage,wesettheβ						to0.1
tingaimstobalancethemodel’sabilitytoadaptto
andcontinuedtrainingfortwoadditionalepochs.
newtrainingsignalswhilemaintainingthestability
Specifically,	During		inference,	we	use	the vllm
-------------	------	---	----------	---	---	--------	---	---	---	---	---	---	---
ofthelearnedbehaviorsfromthereferencemodel.
							Subsequently,		we	trained	five	variants	by ad-
---	---	---	---	---	---	---	-------------	---	---	-------	----	--------	------
6https://github.com/tatsu-lab/stanford_alpaca/
tree/main							justingβ	from0.1to0.5onthemodelpreviously
---------	---	---	---	---	---	---	--------	--------------------------------	---	---	---	---	---

5 answers,	write	at least 5 answers.
Table6: PromptforShort-formQA.
72.0			48.0		23.0
----	---	-----	----	-----	----	-----
68.0			46.0		22.0
64.0			44.0		21.0
60.0		REC.	42.0	REC.	20.0	REC.
56.0		PREC.	40.0	PREC.	19.0	PREC.
2,000 4,000 6,000 8,000 2,000 4,000 6,000 8,000 2,000 4,000 6,000 8,000
		DataSize		DataSize		DataSize
----	---	--------	----	--------	----	--------
76.0			54.0		25.0
72.0			52.0		24.0
68.0			50.0		23.0
64.0			48.0		22.0
		REC.		REC.		REC.
60.0			46.0		21.0
		PREC.		PREC.		PREC.
2,000 4,000 6,000 8,000 2,000 4,000 6,000 8,000 2,000 4,000 6,000 8,000
		DataSize		DataSize		DataSize
---	---	--------	---	--------	---	--------
Figure7: Ablationstudyonsynthetictrainingdatasize: TheupperpartofthefigurecorrespondstotheGrounding
GuidedGenerationtrainingstage, whilethebottompartrepresentstheWeak-to-StrongContrastiveAlignment
trainingstage. Fromlefttoright,theresultsarepresentedforASQA,ELI5,andQAMPARI,respectively. REC.
indicatesCitationRecallandPREC. denotesCitationPrecision. Thex-axisrepresentsthequantityofautomatically
generateddata. Itisobservedthatasthevolumeofautomaticallygenerateddataincreases,thereisaconsistent
improvementinbothcitationrecallandprecisionacrossthetwotrainingstages.
trainedwithG3 toexploretheimpactofthehyper- adheretotheevaluationframeworkestablishedin
parameter β on attribution quality. We evaluated (Gaoetal.,2023b). Forlong-formQAdatasetslike
thesevariantsontheASQAandELI5datasets,and ASQAandELI5,wealsoreportmetricsrelatedto
theexperimentalresultsareshowninFigure9. Fluency,ROUGE-L,andaverageresponselength.
The experimental results indicate that as β in- Specifically,weuseMAUVE(Pillutlaetal.,2021)
creases, the model’s performance on attribution toevaluatethefluencyofthemodelresponse. For
graduallydecreases. Thisobservationsuggeststhat datasetslikeQAMPARI,whereanswersarecom-
thefirststageofG3 mightintroduceanoticeable posed of concatenated entities, we calculate the
inconsistencybetweengroundingandattribution. averagenumberofpredictedentities.
Withhigherβvalues,themodelstrugglestoescape
theconstraintsofinconsistentattributedanswers,
leading to	a reduction	in attribution	quality as	β
----------	-----------	--------------	----------	---	---	---
increases.
F FullResults
Wepresentthecomprehensiveresultsofourexperi-
mentsinTables10,11,and12.		Beyondtheevalua-
-------------------------	---	----------------	---	---	---	---
tionmetricsrelatedtoCorrectnessandCitation,we

context.	Write	a response	that	appropriately	completes		the request.

guide	answer	generation	and	cite the	sources properly.
### Input:Question:		{Question}		Documents:	{Documents}

			Table7:	InstructionFormatforFRONTonLong-formQA.
Below is an instruction that describes a task, paired with an input that provides further
context.	Write	a response	that	appropriately	completes		the request.
--------	-----	----------	----	-------------	---------	---	------------	---	---	---	---

for each	answer.	Separate	answers	by	commas.
### Input:Question:		{Question}		Documents:	{Documents}

			Table8:	InstructionFormatforFRONTonShort-formQA.
	Hyper-parameters			Value
	Top-p			0.95
	Temperature			0.2
	Max-length			2048
Table9:	Hyper-parametersettingsininference.
							72		56
						)%(1FnoitatiC	71		)%(1FnoitatiC 55
Mistral-7B
			VANILLA-SFT(Mistral-7B)				70		54
---	---	---	-----------------------	---	---	---	---	---	---	---	---
			FRONT(Mistral-7B)				69		53
60
							68		52
----------------	---	---	---	---	---	---	--------------	-----------	---	--------------	-----------
)%(1FnoitatiC 50							67	FRONT-7B	51		FRONT-7B
40							0.1 0.2	0.3 0.4 0.5		0.1 0.2	0.3 0.4 0.5
30							βonASQAdataset			βonELI5dataset
20
						Figure	9: Ablation	on hyperparameter		β	in Weak-to-
---	---	---	---	---	---	--------------------------------------------	-----------	-----------------	---	---	-----------
10						StrongContrastiveAlignmentstageonASQAandELI5
0
	ASQA		ELI5	QAMPARI
---	----	---	----	-------	---	---	---	---	---	---	---
Figure8:Ablationstudyonmodelarchitecture:Wesub-
stitutedthefoundationmodelinFRONTwithMistral-7B
andcomparedtheexperimentalresultsofmodelsunder
the same	foundation	model	using	in-context	learning
--------	----------	-----	-----	----------	--------	---	---	---	---	---	---
andthosedirectlysupervisedfine-tunedonourautomat-
icallygenerateddata.		Theexperimentsdemonstratethat
--------------------	---------	-----------------------------	---	-------	----------	---	---	---	---	---	---
by replacing	different	foundation		models,	our frame-
workstillmaintainsitsgeneralizability.

		Fluency	Correct.	Citation
ModelType	ModelSize
		(MAUVE)	(EMRec.)	Rec. Prec.	F1 ROUGE-L	Length
Prompting-based
ChatGPT	-	73.41	40.37	72.81 69.69	71.22 37.92	39.24
------------	---	-----	-----	-----------	-----------	-----
	7B	79.90	24.32	17.24 17.87	17.55 29.38	42.29
LLaMA-2	13B	87.08	27.99	16.45 19.04	17.65 31.41	39.25
	70B	69.28	31.53	44.18 44.79	44.48 31.53	26.86
	7B	66.78	29.93	55.99 51.66	53.74 32.93	26.18
LLaMA-2-Chat	13B	66.14	34.39	37.15 38.17	37.65 35.13	33.68
	70B	86.60	41.24	60.19 61.16	60.67 37.01	47.09
	7B	86.92	38.34	48.37 44.63	46.42 35.95	63.90
Vicuna-v1.5
	13B	66.11	35.20	51.92 53.40	52.65 35.74	38.57
---	---	-----	-----	-----------	-----------	-----
	7B	82.37	29.46	23.12 25.45	24.23 31.67	37.17
Mistral
	8×7B	83.30	36.30	32.72 34.49	33.58 35.05	38.47
---	----	-----	-----	-----------	-----------	-----
	7B	82.86	38.57	64.90 59.67	62.18 36.21	45.26
Mistral-Instruct
	8×7B	94.77	44.11	61.80 63.27	62.53 38.54	58.83
---	----	-----	-----	-----------	-----------	-----
Post-hocRetrieval
ChatGPT	-	49.78	37.68	27.11 27.05	27.08 36.64	52.61
------------	----	-----	-----	-----------	-----------	------
	7B	75.56	16.55	13.88 13.86	13.87 26.81	37.50
LLaMA-2	13B	77.91	20.51	20.95 20.94	20.94 29.53	31.37
	70B	75.23	27.58	28.43 28.43	28.43 30.33	29.88
	7B	22.50	14.17	11.33 11.33	11.33 21.17	110.04
LLaMA-2-Chat	13B	64.52	24.43	21.43 21.43	21.43 33.91	41.12
	70B	70.63	29.68	24.51 24.51	24.51 34.17	45.74
	7B	63.87	19.58	16.24 16.24	16.24 33.22	41.80
Vicuna-v1.5	13B	73.83	24.79	24.11 24.11	24.11 34.42	43.54
	7B	86.54	21.17	16.78 16.77	16.77 30.90	42.43
Mistral	8×7B	80.99	36.30	38.37 35.27	36.75 35.05	38.47
	7B	67.97	26.26	17.87 17.85	17.86 33.71	51.56
Mistral-Instruct 8×7B 65.51 33.90 24.57 24.48 24.52 36.20 53.83
Training-based
	7B	74.33	29.96	67.82 66.97	67.39 35.70	29.83
---	---	-----	-----	-----------	-----------	-----
Self-RAG
	13B	71.59	31.66	71.26 70.35	70.80 36.01	27.03
---	---	-----	-----	-----------	-----------	-----
	7B	76.66	40.32	67.67 63.67	65.61 38.32	62.00
VANILLA-SFT
	13B	84.36	40.85	71.49 66.21	68.75 38.22	58.82
---	---	-----	-----	-----------	-----------	-----
	7B	81.88	40.84	77.70 69.89	73.59 36.95	53.93
FRONT
	13B	76.11	41.51	78.44 73.66	75.95 38.63	57.56
---	---	--------	----------------	-----------	-----------	-----
		Table10:	ASQAfullresults.

		Fluency	Correct.	Citation
ModelType	ModelSize
		(MAUVE)	(Claim) Rec.	Prec. F1	ROUGE-L	Length
Prompting-based
ChatGPT	-	44.65	12.47 49.44	47.05 48.22	20.64	90.2
------------	---	-----	-----------	-----------	-----	------
	7B	63.72	4.53 3.92	5.38 4.54	18.27	103.36
LLaMA-2	13B	62.19	7.77 8.49	8.43 8.46	19.95	88.23
	70B	53.39	10.43 23.75	22.43 23.07	20.43	93.84
	7B	32.80	12.47 19.90	15.48 17.41	20.88	96.42
LLaMA-2-Chat	13B	29.08	13.83 16.50	16.09 16.29	21.04	94.32
	70B	33.69	13.30 36.63	36.63 36.63	21.29	117.84
	7B	31.45	12.30 29.81	22.45 25.61	21.36	105.68
Vicuna-v1.5
	13B	37.41	14.33 31.15	28.99 30.03	21.74	98.23
---	---	-----	-----------	-----------	-----	-----
	7B	56.62	8.47 16.04	16.32 16.18	20.46	93.80
Mistral
	8×7B	61.83	10.43 26.11	25.09 25.59	20.66	93.59
---	----	-----	-----------	-----------	-----	-----
	7B	32.74	11.07 49.25	42.69 45.74	20.75	98.28
Mistral-Instruct
	8×7B	38.51	13.93 49.28	48.34 48.81	21.34	113.71
---	----	-----	-----------	-----------	-----	------
Post-hocRetrieval
ChatGPT	-	22.79	18.77 14.55	14.55 14.55	22.28	106.83
------------	---	-----	-----------	-----------	-----	------
	7B	72.80	7.23 6.84	6.84 6.84	19.14	88.19
LLaMA-2	13B	53.21	10.33 9.61	9.61 9.61	20.63	90.44
	70B	58.97	11.10 10.27	10.26 10.26	20.41	77.85
	7B	22.50	14.17 11.33	11.33 11.33	21.17	110.04
LLaMA-2-Chat	13B	30.36	14.93 12.10	12.10 12.10	21.82	109.79
	70B	37.87	16.03 12.93	12.93 12.93	21.57	99.94
	7B	30.88	11.83 10.91	10.91 10.91	21.66	99.03
Vicuna-v1.5
	13B	32.59	15.20 14.06	14.06 14.05	14.05	108.16
---	---	-----	-----------	-----------	-----	------
	7B	52.45	10.47 8.64	8.64 8.64	20.48	90.17
Mistral
	8×7B	48.39	13.57 11.62	11.62 11.62	21.43	91.97
---	----	-----	-----------	-----------	-----	------
	7B	27.41	17.07 13.20	13.20 13.20	21.52	106.93
Mistral-Instruct
	8×7B	27.60	17.37 15.68	15.68 15.68	21.66	95.21
---	----	-----	-----------	-----------	-----	-----
Training-based
	7B	30.98	6.90 22.34	32.40 26.45	16.48	41.66
---	---	-----	----------	-----------	-----	-----
Self-RAG
	13B	32.04	6.07 30.46	40.20 34.66	15.23	38.19
---	---	-----	----------	-----------	-----	-----
	7B	44.12	9.63 42.30	40.06 41.15	20.58	80.43
VANILLA-SFT
	13B	46.33	10.27 46.75	44.47 45.58	20.56	84.01
---	---	-----	-----------	-----------	-----	-----
	7B	36.90	9.18 58.60	55.33 56.92	19.09	74.06
FRONT
	13B	34.37	9.32 60.31	59.21 59.75	19.66	75.14
---	---	--------	----------------	-----------	-----	-----
		Table11:	ELI5fullresults.

		Correctness	Citation
ModelType	ModelSize
		Rec.-5 Prec.	Rec. Prec.	F1 NumPred.
Prompting-based
ChatGPT	-	20.28 19.84	19.06 22.03	20.44 4.71
------------	---	-----------	-----------	----------
	7B	12.56 11.32	6.03 6.35	6.19 7.02
LLaMA-2	13B	18.00 12.39	5.45 5.74	5.59 11.31
	70B	18.50 14.79	10.10 10.50	10.30 8.31
	7B	17.96 19.74	9.58 9.68	9.63 4.73
LLaMA-2-Chat	13B	21.34 18.86	8.94 9.06	9.00 6.51
	70B	22.62 18.04	13.49 13.98	13.73 7.44
	7B	14.22 14.74	11.26 11.64	11.45 5.87
Vicuna-v1.5
	13B	22.06 19.60	13.04 13.74	13.38 7.62
---	---	-----------	-----------	----------
	7B	16.96 15.98	7.50 7.76	7.63 6.29
Mistral
	8×7B	18.18 15.63	9.72 10.20	9.95 6.63
---	----	-----------	-----------	----------
	7B	17.52 21.29	17.56 18.53	18.03 4.54
Mistral-Instruct
	8×7B	20.12 19.64	19.27 20.38	19.81 5.32
---	----	-----------	-----------	----------
Post-hocRetrieval
ChatGPT	-	25.14 22.85	12.29 12.29	12.29 5.46
------------	---	-----------	-----------	----------
	7B	6.48 5.11	5.05 5.05	5.05 6.55
LLaMA-2	13B	9.88 7.17	5.20 5.20	5.20 6.98
	70B	14.44 12.44	7.49 7.49	7.49 7.41
	7B	12.94 10.89	7.76 7.76	7.76 5.99
LLaMA-2-Chat	13B	15.72 12.23	7.87 7.87	7.87 6.32
	70B	17.90 14.45	9.05 9.05	9.05 6.05
	7B	12.04 9.71	6.69 6.69	6.69 7.10
Vicuna-v1.5
	13B	14.78 11.47	8.50 8.50	8.50 6.67
---	---	-----------	---------	---------
	7B	9.94 7.90	6.00 6.00	6.00 7.38
Mistral
	8×7B	13.92 12.08	6.70 6.70	6.70 6.58
---	----	-----------	---------	---------
	7B	15.80 12.15	8.34 8.34	8.34 7.01
Mistral-Instruct
	8×7B	24.16 18.28	9.78 9.78	9.78 7.37
---	----	-----------	---------	---------
Training-based
	7B	2.34 1.98	10.53 18.80	13.50 3.49
---	---	---------	-----------	----------
Self-RAG
	13B	1.90 1.33	12.79 20.90	15.86 3.08
---	---	-----------	-----------	----------
	7B	12.86 21.09	21.35 21.36	21.35 7.49
VANILLA-SFT
	13B	12.68 22.80	23.64 23.71	23.67 3.14
---	---	-----------	-----------	----------
	7B	11.50 21.38	24.74 24.84	24.79 3.08
FRONT
	13B	11.94 22.61	24.86 25.39	25.12 3.17
---	--------	-------------------	-----------	----------
	Table12:	QAMPARIfullresults.

Datasets:

luoojason
/

muddle-eval-bundle-md

Instruction:

Response:

Instruction:

Response: