What Makes Good In-Context Demonstrations for Code Intelligence Tasks with LLMs?

Pre-trained models of source code have gained widespread popularity in many code intelligence tasks. Recently, with the scaling of the model and corpus size, large language models have shown the ability of in-context learning (ICL). ICL employs task instructions and a few examples as demonstrations,...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:IEEE/ACM International Conference on Automated Software Engineering : [proceedings] S. 761 - 773
Hauptverfasser: Gao, Shuzheng, Wen, Xin-Cheng, Gao, Cuiyun, Wang, Wenxuan, Zhang, Hongyu, Lyu, Michael R.
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: IEEE 11.09.2023
Schlagworte:
ISSN:2643-1572
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Abstract Pre-trained models of source code have gained widespread popularity in many code intelligence tasks. Recently, with the scaling of the model and corpus size, large language models have shown the ability of in-context learning (ICL). ICL employs task instructions and a few examples as demonstrations, and then inputs the demonstrations to the language models for making predictions. This new learning paradigm is training-free and has shown impressive performance in various natural language processing and code intelligence tasks. However, the performance of ICL heavily relies on the quality of demonstrations, e.g., the selected examples. It is important to systematically investigate how to construct a good demonstration for code-related tasks. In this paper, we empirically explore the impact of three key factors on the performance of ICL in code intelligence tasks: the selection, order, and number of demonstration examples. We conduct extensive experiments on three code intelligence tasks including code summarization, bug fixing, and program synthesis. Our experimental results demonstrate that all the above three factors dramatically impact the performance of ICL in code intelligence tasks. Additionally, we summarize our findings and provide takeaway suggestions on how to construct effective demonstrations, taking into account these three perspectives. We also show that a carefully-designed demonstration based on our findings can lead to substantial improvements over widely-used demonstration construction methods, e.g., improving BLEU-4, EM, and EM by at least 9.90%, 175.96%, and 50.81% on code summarization, bug fixing, and program synthesis, respectively.
AbstractList Pre-trained models of source code have gained widespread popularity in many code intelligence tasks. Recently, with the scaling of the model and corpus size, large language models have shown the ability of in-context learning (ICL). ICL employs task instructions and a few examples as demonstrations, and then inputs the demonstrations to the language models for making predictions. This new learning paradigm is training-free and has shown impressive performance in various natural language processing and code intelligence tasks. However, the performance of ICL heavily relies on the quality of demonstrations, e.g., the selected examples. It is important to systematically investigate how to construct a good demonstration for code-related tasks. In this paper, we empirically explore the impact of three key factors on the performance of ICL in code intelligence tasks: the selection, order, and number of demonstration examples. We conduct extensive experiments on three code intelligence tasks including code summarization, bug fixing, and program synthesis. Our experimental results demonstrate that all the above three factors dramatically impact the performance of ICL in code intelligence tasks. Additionally, we summarize our findings and provide takeaway suggestions on how to construct effective demonstrations, taking into account these three perspectives. We also show that a carefully-designed demonstration based on our findings can lead to substantial improvements over widely-used demonstration construction methods, e.g., improving BLEU-4, EM, and EM by at least 9.90%, 175.96%, and 50.81% on code summarization, bug fixing, and program synthesis, respectively.
Author Wen, Xin-Cheng
Gao, Cuiyun
Gao, Shuzheng
Lyu, Michael R.
Wang, Wenxuan
Zhang, Hongyu
Author_xml – sequence: 1
  givenname: Shuzheng
  surname: Gao
  fullname: Gao, Shuzheng
  email: szgao98@gmail.com
  organization: School of Computer Science and Technology, Harbin Institute of Technology,Shenzhen,China
– sequence: 2
  givenname: Xin-Cheng
  surname: Wen
  fullname: Wen, Xin-Cheng
  email: xiamenwxc@foxmail.com
  organization: School of Computer Science and Technology, Harbin Institute of Technology,Shenzhen,China
– sequence: 3
  givenname: Cuiyun
  surname: Gao
  fullname: Gao, Cuiyun
  email: gaocuiyun@hit.edu.cn
  organization: School of Computer Science and Technology, Harbin Institute of Technology,Shenzhen,China
– sequence: 4
  givenname: Wenxuan
  surname: Wang
  fullname: Wang, Wenxuan
  email: wxwang@cse.cuhk.edu.hk
  organization: The Chinese University of Hong Kong,Department of Computer Science and Engineering,China
– sequence: 5
  givenname: Hongyu
  surname: Zhang
  fullname: Zhang, Hongyu
  email: hyzhang@cqu.edu.cn
  organization: School of Big Data and Software Engineering, Chongqing University,China
– sequence: 6
  givenname: Michael R.
  surname: Lyu
  fullname: Lyu, Michael R.
  email: lyu@cse.cuhk.edu.hk
  organization: The Chinese University of Hong Kong,Department of Computer Science and Engineering,China
BookMark eNotj9FKwzAYhaMouM09gV7kBVqTP23TXMmocw46FJx4OZLmj6vrGmkC6ttb1KuPczgczpmSs973SMgVZynnTN0snpd5AaBSYCBSxkbvhMyVVKXImQCliuyUTKDIRMJzCRdkGsI7Y_ko5IQ8ve51pBt9wEBX3lu67pPK9xG_Ir3Do-9DHHRsR1LnB1p5i2MkYte1b9g3SLc6HAL9bOOe1vUm3F6Sc6e7gPN_zsjL_XJbPST142pdLepEQ5nFRJaA-neeyQyXhjlnnFJNDgiMNU421vLSNdqUFrjWFoVsSmm5Rq6MUGJGrv96W0TcfQztUQ_fO85g_A1K_ACrmVG6
CODEN IEEPAD
ContentType Conference Proceeding
DBID 6IE
6IL
CBEJK
RIE
RIL
DOI 10.1109/ASE56229.2023.00109
DatabaseName IEEE Electronic Library (IEL) Conference Proceedings
IEEE Proceedings Order Plan All Online (POP All Online) 1998-present by volume
IEEE Xplore All Conference Proceedings
IEEE Electronic Library (IEL)
IEEE Proceedings Order Plans (POP All) 1998-Present
DatabaseTitleList
Database_xml – sequence: 1
  dbid: RIE
  name: IEEE Electronic Library (IEL)
  url: https://ieeexplore.ieee.org/
  sourceTypes: Publisher
DeliveryMethod fulltext_linktorsrc
Discipline Computer Science
EISBN 9798350329964
EISSN 2643-1572
EndPage 773
ExternalDocumentID 10298329
Genre orig-research
GrantInformation_xml – fundername: Natural Science Foundation of Guangdong Province
  grantid: 2023A1515011959
  funderid: 10.13039/501100003453
– fundername: National Natural Science Foundation of China
  grantid: 62002084
  funderid: 10.13039/501100001809
GroupedDBID 6IE
6IF
6IH
6IK
6IL
6IM
6IN
6J9
AAJGR
AAWTH
ABLEC
ACREN
ADYOE
ADZIZ
AFYQB
ALMA_UNASSIGNED_HOLDINGS
AMTXH
BEFXN
BFFAM
BGNUA
BKEBE
BPEOZ
CBEJK
CHZPO
IEGSK
IPLJI
M43
OCL
RIE
RIL
ID FETCH-LOGICAL-a284t-782ea03299b4b17b0ffbf99c52e200cf7cdd18fcab8d21aade37c87d1ae19b393
IEDL.DBID RIE
ISICitedReferencesCount 56
ISICitedReferencesURI http://www.webofscience.com/api/gateway?GWVersion=2&SrcApp=Summon&SrcAuth=ProQuest&DestLinkType=CitingArticles&DestApp=WOS_CPL&KeyUT=001103357200061&url=https%3A%2F%2Fcvtisr.summon.serialssolutions.com%2F%23%21%2Fsearch%3Fho%3Df%26include.ft.matches%3Dt%26l%3Dnull%26q%3D
IngestDate Wed Aug 27 02:32:41 EDT 2025
IsPeerReviewed false
IsScholarly true
Language English
LinkModel DirectLink
MergedId FETCHMERGED-LOGICAL-a284t-782ea03299b4b17b0ffbf99c52e200cf7cdd18fcab8d21aade37c87d1ae19b393
PageCount 13
ParticipantIDs ieee_primary_10298329
PublicationCentury 2000
PublicationDate 2023-Sept.-11
PublicationDateYYYYMMDD 2023-09-11
PublicationDate_xml – month: 09
  year: 2023
  text: 2023-Sept.-11
  day: 11
PublicationDecade 2020
PublicationTitle IEEE/ACM International Conference on Automated Software Engineering : [proceedings]
PublicationTitleAbbrev ASE
PublicationYear 2023
Publisher IEEE
Publisher_xml – name: IEEE
SSID ssj0051577
ssib057256115
Score 2.5887785
Snippet Pre-trained models of source code have gained widespread popularity in many code intelligence tasks. Recently, with the scaling of the model and corpus size,...
SourceID ieee
SourceType Publisher
StartPage 761
SubjectTerms Codes
Computer bugs
Natural language processing
Predictive models
Software engineering
Source coding
Task analysis
Title What Makes Good In-Context Demonstrations for Code Intelligence Tasks with LLMs?
URI https://ieeexplore.ieee.org/document/10298329
WOSCitedRecordID wos001103357200061&url=https%3A%2F%2Fcvtisr.summon.serialssolutions.com%2F%23%21%2Fsearch%3Fho%3Df%26include.ft.matches%3Dt%26l%3Dnull%26q%3D
hasFullText 1
inHoldings 1
isFullTextHit
isPrint
link http://cvtisr.summon.serialssolutions.com/2.0.0/link/0/eLvHCXMwlV1LSwMxEA62ePBUHxXf5OA1utmHszmJ1FYLbSlYobeSTSYgxd3S3Yo_3yTd1l48eAs5hZkM30wy3zeE3BoJsYgxZaghY3GEyGSMCeMoExNolaL0kvkDGI3S6VSMa7K658Igom8-wzu39H_5ulAr91RmIzwU9gaKBmkAwJqstbk8CVjw5nyb-1qcBqhlhngg7p_euhbqQ8dNCZ2oqW9B3Bmo4vGk1_rnSQ5J-5eZR8dbzDkie5gfk9ZmNAOtI_WEjJ0kNx3KOZb0pSg07efMC1F9V_QZP11SuHZ9SW3WSjuFRtrfUeekE1nOS-peaelgMCwf2-S91510Xlk9PIFJizgVs8iPMrBHFFmcccgCYzIjhEpCtIGhDCiteWqUzFIdcik1RqBS0FwiF1kkolPSzIsczwhNFIIrZAOldAy2pg0lyAdjS0EhjE0Iz0nbWWi2WOtjzDbGufhj_5IcOCe4rgvOr0izWq7wmuyrr-qjXN54r_4AaTyizA
linkProvider IEEE
linkToHtml http://cvtisr.summon.serialssolutions.com/2.0.0/link/0/eLvHCXMwlV1NSwMxEA1aBT3Vj4rf5uA1utnNms1JpFpb3JaCFXor2WQCUtyV7lb8-SbptvbiwVvIKcxkeDPJvDcIXRvJmWCQENA8IywCIJJBTCjI2ARaJSC9ZH7KB4NkPBbDmqzuuTAA4JvP4MYt_V--LtTcPZXZCA-FvYFiE23FjIV0QddaXp-YW_imdJX9WqTmvBYaooG4fXh9smAfOnZK6GRNfRPi2kgVjyid5j_Psodav9w8PFyhzj7agPwANZfDGXAdq4do6ES5cV9OocTPRaFxLydeiuq7wo_w4dLChfNLbPNW3C404N6aPiceyXJaYvdOi9O0X9630FvnadTuknp8ApEWcypisR9kYI8oMpZRngXGZEYIFYdgQ0MZrrSmiVEyS3RIpdQQcZVwTSVQkUUiOkKNvMjhGOFYAXelbKCUZtxWtaHk8s7YYlAIY1PCE9RyFpp8LhQyJkvjnP6xf4V2uqN-Okl7g5cztOsc4nowKD1HjWo2hwu0rb6q93J26T38AxJpphM
openUrl ctx_ver=Z39.88-2004&ctx_enc=info%3Aofi%2Fenc%3AUTF-8&rfr_id=info%3Asid%2Fsummon.serialssolutions.com&rft_val_fmt=info%3Aofi%2Ffmt%3Akev%3Amtx%3Abook&rft.genre=proceeding&rft.title=IEEE%2FACM+International+Conference+on+Automated+Software+Engineering+%3A+%5Bproceedings%5D&rft.atitle=What+Makes+Good+In-Context+Demonstrations+for+Code+Intelligence+Tasks+with+LLMs%3F&rft.au=Gao%2C+Shuzheng&rft.au=Wen%2C+Xin-Cheng&rft.au=Gao%2C+Cuiyun&rft.au=Wang%2C+Wenxuan&rft.date=2023-09-11&rft.pub=IEEE&rft.eissn=2643-1572&rft.spage=761&rft.epage=773&rft_id=info:doi/10.1109%2FASE56229.2023.00109&rft.externalDocID=10298329