| | | 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- Automatic Data Generation Pipeline           Data Collection        Answer Generation           Data Filtering Question: Why do bagels have holes in the middle？ Bagel holes ensures even baking (...) a area | | | | | | | 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. Two-Stage Training Recipe           Grounding Guided Generation        Consistency-Aware Alignment           Answer Generation | 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) | (3) | | ------------------ | --- | ---------------------------- | --- | --- | --- | --- | --- | ----- | ------ | ----- | --- | --- | | 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 groundedquotesasfine-grainedverification. Inre- whereπ ref representsthereferencemodel,initial- sponse,weproposeaconsistency-awarealignment izedfromG3. Thehyper-parameterβ modulates (CCA) stage specifically aimed at enhancing the the divergence between the distribution from the consistencybetweenthegroundingprocessandthe policy model and the reference model. τ is the w generationprocess. consistentanswer,whileτ istheinconsistentone. l | | 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 | exact | | | | | 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. | | | | | | 4.4 ImplementDetails | 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 leading open-source models and two variants of Authenticity Helpfulness | 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. | | | | | | | ------- | ------ | -------- | --------- | ----- | ------- | -------------- | --- | --- | --- | --- | --- | Acknowledgements qa-basedfactualconsistencyevaluationforsumma- | | | | | | | | 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, WeihuaPeng,XiaochengFeng,BingQin,andTing servingwithpagedattention. InProceedingsofthe Liu. 2023. A survey on hallucination in large lan- ACMSIGOPS29thSymposiumonOperatingSystems | 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, NeurIPS2022,NewOrleans,LA,USA,November28 Doshi,RenelitoDelosSantos,TojuDuke,JohnnySo- | -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. | | | | | | | ---------------- | --- | -------- | ------ | ------- | --- | ------ | --------------- | --- | --- | --- | --- | --- | Kun Zhu, Xiaocheng Feng, Xiyuan Du, Yuxuan Gu, process, we retained 5,667 and 2,431 queries, re- WeijiangYu,HaotianWang,QianglongChen,Zheng spectively. Additionally,wecalculatedtheaverage | 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. | | | | | | | ------------- | --- | ----------------- | ---- | --------- | ----------------- | --- | --- | --- | --- | --- | Fluency Weevaluatethefluencyofthegenerated framework(Kwonetal.,2023)forefficientinfer- responseusingMAUVE(Pillutlaetal.,2021). No- ence. Thehyperparametersaresetasillustratedin | 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 | | | | | | | --------- | --- | --- | --- | --- | --- | --- | -------- | -------------------------------- | --- | --- | --- | --- | --- | Instruction: Write an accurate, engaging, and concise answer for the given question using only the provided search results (some of which might be irrelevant) and cite them properly. Use an unbiased and journalistic tone. Always cite for any factual claim. When citing several search results, use [1][2][3]. Cite at least one document and at most three documents in each sentence. If multiple documents support the sentence, only cite a minimum sufficient subset of the documents. Table5: PromptforLong-formQA. Instruction: Provide a list of accurate answers for the given question using only the provided search results (some of which might be irrelevant) and cite them properly. Always cite one and only one document for each answer. Separate answers by commas. For questions that have more than | 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 Below is an instruction that describes a task, paired with an input that provides further | context. | Write | a response | that | appropriately | completes | | the request. | | | | | | -------- | ----- | ---------- | ---- | ------------- | --------- | --- | ------------ | --- | --- | --- | --- | ### Instruction: Extract the relevant content from the provided documents and then use the extracted content to | guide | answer | generation | and | cite the | sources properly. | | | | | | | | ------------------- | ------ | ---------- | --- | ---------- | ----------------- | --- | --- | --- | --- | --- | --- | | ### Input:Question: | | {Question} | | Documents: | {Documents} | | | | | | | ### Response: | | | | 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. | | | | | | -------- | ----- | ---------- | ---- | ------------- | --------- | --- | ------------ | --- | --- | --- | --- | ### Instruction: Extract the relevant content from the provided documents and then use the extracted content to provide a list of accurate answers for the given question. Always cite one and only one document | for each | answer. | Separate | answers | by | commas. | | | | | | | | ------------------- | ------- | ---------- | ------- | ---------- | ----------- | --- | --- | --- | --- | --- | --- | | ### Input:Question: | | {Question} | | Documents: | {Documents} | | | | | | | ### Response: | | | | 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. | | |