Paper deep dive

The Order Is the Guarantee: Verifier-Budgeted Code Deletion with Static-First Learned Proposals

Ruitong Li, Binjie Guo, Aisheng Mo, Guowei Su, Han Wang, Jie Li, Ru Zhang

Year: 2026Venue: arXiv preprintArea: cs.SEType: PreprintEmbeddings: 50

Intelligence

Status: not_run | Model: - | Prompt: - | Confidence: 0%

Entities (0)

No extracted entities yet.

Relation Signals (0)

No relation signals yet.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Frontier coding models now match or exceed strong human reference points on programming benchmarks, yet benchmark success does not imply maintainable software. Prompt-driven "vibe coding" is additive: new branches, guards, and fallbacks accumulate faster than obsolete logic is removed. We study the inverse problem-how an Al system should remove code when execution-verification capacity is finite. We formulate redundant-code reduction as proposal scheduling: a ranker orders single-statement deletion candidates, an execution suite accepts the first candidate that passes, and a budget bounds how many candidates may be tested. Our central observation is that candidate order, not model confidence, is the control surface a deployment can reason about. DELSCOUT instantiates two schedules. Given representative target-domain validation, a five-slot budget spends three slots on deterministic shortest-first candidates and two on complementary learned candidates; across nine MBPP replications with 0.5B, 0.6B, and 8B rankers this raises verified-deletion coverage by 9.5% relative (+6.7 accepted tasks) while consuming slightly fewer verifier calls than the matched static baseline. Without such validation the same rankers can lose coverage under shift, so we instead evaluate the complete static prefix first and append learned candidates only afterwards; for a deterministic verifier this makes coverage and character reduction non-decreasing by construction, at a measured 4.8-62.5% increase in verifier calls. MBPP+ then erases the in-domain advantage, showing that scheduling governs search while the test suite alone governs what "preserving behavior" means. The result is an auditable division of labor: models widen the search for removable code, order bounds the damage a mis-ranked proposal can do, and execution retains authority over every committed deletion.

PDF

Open source PDF →

PDF not stored locally. Use the link above to view on the source site.

Full Text

47,094 characters extracted from source content.

Expand or collapse full text

1 TheOrderIstheGuarantee:Verifier-BudgetedCodeDeletion withStatic-FirstLearnedProposals RuitongLi 1,* BinjieGuo 2,* AishengMo 2 GuoweiSu 2 HanWang 3 JieLi 4 RuZhang 2 1 UniversityofHongKong 2 ZhejiangUniversity 3 DalianUniversityofTechnology 4 IndependentResearcher * Theseauthorscontributedequally. Abstract Frontiercodingmodelsnowmatchorexceedstronghumanreferencepointsonprogramming benchmarks,yetbenchmarksuccessdoesnotimplymaintainablesoftware.Prompt-driven“vibecoding” isadditive:newbranches,guards,andfallbacksaccumulatefasterthanobsoletelogicisremoved.We studytheinverseproblem—howanAIsystemshouldremovecodewhenexecution-verificationcapacity isfinite.Weformulateredundant-codereductionas proposalscheduling :arankerorderssingle- statementdeletioncandidates,anexecutionsuiteacceptsthefirstcandidatethatpasses,andabudget boundshowmanycandidatesmaybetested.Ourcentralobservationisthatcandidate order ,not modelconfidence,isthecontrolsurfaceadeploymentcanreasonabout.DelScoutinstantiatestwo schedules.Givenrepresentativetarget-domainvalidation,afive-slotbudgetspendsthreeslotson deterministicshortest-firstcandidatesandtwooncomplementarylearnedcandidates;acrossnine MBPPreplicationswith0.5B,0.6B,and8Brankersthisraisesverified-deletioncoverageby9.5% relative(+6.7acceptedtasks)whileconsumingslightlyfewerverifiercallsthanthematchedstatic baseline.Withoutsuchvalidationthesamerankerscanlosecoverageundershift,soweinstead evaluatethecompletestaticprefixfirstandappendlearnedcandidatesonlyafterwards;fora deterministicverifierthismakescoverageandcharacterreductionnon-decreasingbyconstruction,ata measured4.8–62.5%increaseinverifiercalls.MBPP+thenerasesthein-domainadvantage,showing thatschedulinggovernssearchwhilethetestsuitealonegovernswhat“preservingbehavior”means. Theresultisanauditabledivisionoflabor:modelswidenthesearchforremovablecode,orderboundsthe damageamis-rankedproposalcando,andexecutionretainsauthorityovereverycommitteddeletion. 1Introduction Largelanguagemodelshavetransformedprogrammingfromtoken-levelcompletionintoend-to-endproblem solving.SystemssuchasAlphaCodeandDeepSeek-R1nowmatchorsurpassstronghumanbaselines oncompetitive-programmingbenchmarks,whilecontemporarycodemodelstackleincreasinglydiverse generationtasksrangingfromrepository-leveleditingtofull-featureimplementation(Lietal.,2022;DeepSeek- AI,2025;Huietal.,2024;Jimenezetal.,2024).Thisprogress,however,shiftsthebottleneck.Producingyet anotherworkingimplementationisbecomingcheap;establishingthateverygeneratedbranch,helperfunction, importstatement,andcompatibilitylayerremainsgenuinelynecessaryisnot. Thisimbalanceismostevidentinproject-scale“vibecoding,”wheredevelopersprompt,patch,and regenerateuntiltestspass.Eachiterationtendstoaddalocalfixwhilepreservingearlierscaffolding,so abandonedalternativespersistandresponsibilitiesbecomeduplicatedacrossedits.Theprogramremains functional,butredundantcodeexpandsthereviewsurface,obscuresinvariants,andsolidifiesintotechnical 2 Validatedmixture,|π|=5Prefix-preservingaugmentation, |π|≤7 Executionverifier commitfirstpassingdeletion SSSLL LearnedorderM twolearnedslotsreplacestaticslots4–5 Figure1:DelScoutseparatesproposalsfromacceptance.Thevalidatedmixturespendsavalidatedfive-slotbudget onathree-candidatestaticcoreplustwocomplementarylearnedcandidates.prefix-preservingaugmentation keepstheentirefive-candidatestaticrankingandappendslearnedcandidatesonlyafterstaticfailure,soalearned rankingcanneverdisplaceastaticsuccess.SandLdenotestaticandlearnedproposals. debt—asmellthatdevelopersrecognizebutrarelyremove(Romanoetal.,2020;Kruchtenetal.,2012). Currentgenerationandrepairbenchmarksrewardcodethatpassestests;almostnonerewardasystemfor identifyingwhatshouldnolongerexist.Westudythisinverseproblemasverifier-backedcode deletion :proposeastatementtoremove,executetherelevanttests,andcommitonlyifalldeletionspass. Deletionisintrinsicallyasymmetric.Ausefulproposalreducesmaintenanceburden,whereasawrong deletionsilentlyremovesrequiredbehavior—whichiswhyprogramreducershavealwaysbeenorganized aroundanexecutableoracle(ZellerandHildebrandt,2002;Regehretal.,2012;MisherghiandSu,2006; Sunetal.,2018).Soundstaticanalysiscapturesthesafehalfofthisproblem:compilersandlintersremove codetheycanproveunreachableorunused(Knoopetal.,1994).Thehardhalfiscodethatisreachableand referencedyetbehaviorallysubsumedbythesurvivingimplementation—aredundantguard,aduplicated normalizationstep,anobsoletefallback.Judgingthoserequirescontextualunderstanding,whichiswherea learnedrankercanhelp—andwhereamodelscoreisemphaticallynotaproof. Onceexecutionservesastheacceptanceauthority,therealdesignvariablebecomestheschedule: whichproposalsafiniteverificationbudgetisallowedtoevaluate.Thisreframingmattersbecauseit clarifieswhatcanandcannotbeguaranteed.Learnedcandidatesmaycomplementastaticranking,butunder afixedbudgettheycanalsodisplacestaticcandidatesthatwouldhavepassedunderunknowndistributional shift.Ordering,bycontrast,isfreeandfullycontrollable.Oursystem,DelScout(Figure1),letsthemodel ordercandidateswhiletheexecutionsuiteretainsvetopower,andselectsbetweentwoschedulesbasedon theevidenceactuallyavailable: •Whenrepresentativetarget-domainvalidationexists,the validatedmixture allocatesthreeof fiveverificationslotstodeterministicshortest-firstcandidatesandtheremainingtwotothehighest- rankedlearnedcandidatesabsentfromthatcore.Validation,notintuition,licensesthisexchange. •Whensuchvalidationdoesnotexist, prefix-preservingaugmentation evaluatesthe complete five-candidatestaticrankingfirstandappendsatmosttwolearnedcandidatesonlyifallfivestatic candidatesfail.Becausealearnedproposalneveroccupiesastaticslot,augmentationcanaddan accepteddeletionbutcanneverrevokeone.Theguaranteefollowsfromordering,notfrom calibration. Ourcontributionsareasfollows.(i)Weformulateredundant-codereductionasproposalschedulingunder afiniteexecution-verificationbudget,cleanlyseparatinglearnedsearchfromtheacceptanceboundary.(i) Weintroduceatarget-validatedstatic–learnedmixtureandaprefix-preservingaugmentationpolicywhose non-decreasingcoverageandcharacterreductionfollowfromcandidateorderalone.(i)Wereportnine replicationsacross0.5B–8Brankers,objectivecontrolsthatruleoutaprivilegedtrainingloss,budgetand verifier-costaccounting,cross-benchmarkshiftanalysis,andstronger-verifierteststhatlocatepreciselywhere themethodstopstransferring. SSSSSLL allfivestaticslotsretained;learnedtailisadditional Executionverifier commitfirstpassingdeletion Programx ASTcandidatesC(x) StaticorderS shortest/zero-reference verified-successscore notargetlabels 3 2RelatedWork Programreductionandrepair.Reductionhasalwayscoupledcandidateremovalwithan executablenotionofpreservation.Deltadebuggingandhierarchicaldeltadebuggingisolatefailure- inducingfragmentsthroughrepeatedtests(ZellerandHildebrandt,2002;MisherghiandSu,2006);C- ReduceandPersesaddspecializedpassesandgrammarguidanceforefficiencyandsyntacticvalidity (Regehretal.,2012;Sunetal.,2018).Generate-and-validaterepairadoptsthesamesplitbetweensearch andacceptance:GenProg,Prophet,Angelix,TBar,SequenceR,andRecodervarytheproposal mechanismwhiletestsjudgepatches(Weimeretal.,2009;Gouesetal.,2012;LongandRinard,2016; Mechtaevetal.,2016;Liuetal.,2019;Chenetal.,2021b;Zhuetal.,2021).Weinherittheverifier- centeredprinciplebutstudyredundancyremovaland,unlikereducersthatiteratetoalocalminimum,the allocationofasmallfixedproposalbudgetacrossheterogeneousrankers. Deadcodeandtechnicaldebt.Compilerdead-codeeliminationandmodernlintersdeletewhatthey canproveunnecessary,suchasunreachableblocksandunusedbindings(Knoopetal.,1994). Soundnessistheirstrengthandtheirceiling:areachable,referencedstatementwhosebehavioris subsumedelsewhereisinvisibletothem.Empiricalworkreportsthatsuchresidueiswidespread,is perceivedasabadsmellthatimpairscomprehension,andisonlyweaklycorrelatedwithwhere maintenanceeffortisactuallyspent(Romanoetal.,2020;Ederetal.,2012;Kruchtenetal.,2012;Fowler, 1999).Ourmeasurementsmaketheboundaryconcrete:onBigCodeBench-Hardthestrongeststatic rankingsucceedsmainlythroughunusedimports(61of89accepteddeletions),whereasthelearnedtailis dominatedbyexpressionsandconditionals.Learningisthereforenotanalternativetosoundanalysis;it extendsthereachablecandidatesetpastthepointwhereproofisavailable,whichisexactlywhy acceptancemuststaywithexecution. Neuralcodemodelsandexecutionfeedback.Codemodelshaveprogressedfromrepresentation learningtocompetitivegenerationandrepair(Fengetal.,2020;Guoetal.,2021;Wangetal.,2021, 2023;Friedetal.,2023;Lietal.,2023;Huietal.,2024;Lietal.,2022;DeepSeek-AI,2025),and adaptationstudiesshowthatmodestfine-tuningcanreprioritizeusefuledits(Silvaetal.,2024;Huetal., 2022;XiaandZhang,2022).Executionfeedbackhasbeenusedasreward,verification,testsynthesis,and iterativerepair(Leetal.,2022;Nietal.,2023;Chenetal.,2023,2024).HumanEvalandMBPP establishedfunctionalevaluation,whileEvalPlus,DS-1000,BigCodeBench,SWE-bench,andautomated testgenerationexposeweaktests,libraryinteractions,andrepository-scalebehavior(Chenetal.,2021a; Austinetal.,2021;Liuetal.,2023;Laietal.,2023;Zhuoetal.,2025;Jimenezetal.,2024;Lukasczyketal., 2022).Thesesystemsoverwhelminglyoptimizewhattoaddorreplace,andwetreateachsuiteasa distinctoperationalspecificationratherthanapooledscore.Ourcontributionistheschedulingview: executiongateseverydeletion,andprefixpreservationboundshowadistribution-shiftedlearnedrankingcan disturbestablishedstaticcoverage. 3ProblemFormulation LetC(x)bethesetofsyntacticallyvalidsingle-statementASTdeletioncandidatesofprogramx,andlet x .AdeterministicverifierVforadomainreturns1ifx compilesandpassesthatdomain’snamedexecutionsuite,and0otherwise.Aproposalpolicyπreturns anorderedduplicate-freelistπ(x)=(c1,...,cm)with cj ∈C(x);theschedulertestscandidatesleftto rightandcommitsthefirstthatpasses.Writingmin ∅ = ∞ ,theacceptedindexis τ(π,x)=min j ≤ m:V(x \ cj)=1 ,(1) 4 soτ=∞meansthepolicycommitsnodeletionandmisthepolicy’sverifier-callboundforthatprogram. Overaneligiblecorpus D (programswhoseunmodifiedsourcepassesthesamesuite),withr(x,c)the fractionofsourcecharactersremovedbyc,wereport Bothsumsrangeonlyoveracceptedtasks,soataskwithoutadeletioncontributesexactlyzeroandno undefinedcandidateisreferenced.Coverageseparatesarankerthatfindsmanysmalldeletionsfromonethat findsfewerbutlargerremovals,whichRedcaptures. Schedules.LetS(x)=(s1,s2,...)beastaticorderandM(x)=(m1,m2,...)alearnedorderover the sameC(x).WriteSk(x)forthefirstkentriesofS(x),andletM A (x)bethefirstBentriesofM(x)that donotlieinthesetA;∥denotesorderedconcatenation.Thetwoschedulesare πmix(x)=S3(x) ∥M S3(x)(x) , | πmix | =5,(4) πaug(x)=S5(x) ∥M S5(x)(x) , |πaug| ≤5+B.(5) De-duplicationisdefinedagainstthestaticcandidatesthepolicyactuallyproposes—S3in(4)and S5 in (5)—soeachpolicyisagenuinelistofdistinctcandidatesandthetwobudgetsare5and5+BwithB ≤ 2. Proposition1(prefixpreservation).FordeterministicVandeveryB ≥ 0,ifτ(S5,x)=j ≤ 5thenτ(πaug,x)=jandthecommittedcandidateisthesamesj;ifτ(S5,x)= ∞ thenτ(πaug,x)is either∞oranindexinthelearnedtail.HenceCov(πaug)≥Cov(S5)and,becauser≥0,also Red(πaug)≥Red(S5).Theproofisimmediatefrom(5):thefirstfiveentriesofπaug(x)areliterallyS5(x), andfirst-successstoppingmeansalearnedcandidateisreachedonlywhenallfivestaticcandidateshave alreadyfailed.ThestatementassumesadeterministicverifierandsaysnothingaboutbehaviorthatVdoes nottest. Proposition2(nofixed-budgetdominance).FixabudgetKandanyschedulethatomitssome sj withj ≤ KfromSK.Thereisaninstanceonwhich sj istheuniquecandidatewithV(x \ sj)= 1:thenτ(SK,x)=j<∞whilethemixedscheduleneverproposes sj andreturnsτ=∞.No nontrivialfixed-Kreplacementthereforedominatesthecompletestaticprefixoneverytarget distribution.Consequentlyaguaranteeunderarbitraryshiftrequiresextraslotsthatpreservetheprefix, whereaskeepingthebudgetfixedrequiresempiricaltarget-domainvalidation.Thetwoschedulesin(4)–(5) areexactlythesetwooptions,andProposition2iswhyweneverpresentthefixed-budgetresultasa universalclaim. 4Method 4.1WhyTwoProposalFamilies AI-generatedredundancyisnotasinglephenomenon. Syntacticremnants includeimports,assignments, orhelperfunctionsleftbehindafterthecodepaththatrequiredthemhaschanged.Control-flow redundancyariseswhensuccessivepromptsaddguardsoralternativebranchesalreadysubsumed bythesurvivingimplementation.Compatibilityscaffoldingcomprisestemporaryfallbackormigration logicthatoutlivestheenvironmentitoriginallyprotected.Allthreecanappearlocallyplausible:the statementusesmeaningfulidentifiersandresemblesnearbycode,eventhoughremovingitpreservesall testedbehavior. 5 Thesethreeformsrewarddifferentinductivebiases,whichiswhywedeliberatelymaintaintwo heterogeneousrankersratherthanasinglescore.Shortest-firstrankingsuitssyntacticremnants,sinceone-line assignmentsandimportscarrylowdeletionrisk.Zero-referencerankingusesidentifierfrequencytosurface weaklyconnectedstatements.Neithercandeterminewhetheraconditionalisbehaviorallysubsumed,because thatjudgmentdependsonsurroundingcontrolflowandtaskintent;alearnedrankersuppliesthatcontextual comparison.Therankersserveascomplementarysearchoperatorsoverdifferentmanifestationsofgenerated debt,notasinterchangeableestimatorsofasingleunderlyingquantity. WerestricttheactionspacetoonecompleteASTstatement.Thisunitcoverstheremnantsabove,maintains aone-to-onecorrespondencebetweenaproposalandasourcespan,andmakeseveryexecutionoutcome attributabletoasingleedit.Multi-statementtransformationscouldremovelargerduplicatedstructures,but theyentanglecandidategenerationwitheditcompositionandmultiplythepossibleexplanationsforafailure. Establishingtheschedulingprincipleinthesingle-statementsettingprovidesacontrolledfoundationfor repository-scaleextensions. 4.2CandidateSpaceandStaticOrders WeparsePythonusingthestandardlibraryASTandenumeratestatementspansforimports,assignments, definitions,classes,control-flowstatements,contextmanagers,andexpressions.Acandidateiskeyedbystart line,endline,andnodetype;deletingitremovesexactlythecoveredlines.Candidatesthatfailtocompile arerejectedandnevercountedassuccesses.Theshortestordersortsbyspanlength,thencharacter length,thensourceposition,thenkey.Thezero-referenceorderfirstprefersstatementswhoseidentifiers occurleastfrequentlyinthefile,thenshorterspans.Botharedeterministicandparameter-free,andwhich oneislocallystrongerisdecidedperdomainwithoutconsultinglearnedtestoutcomes. 4.3LearningVerifiedUtility,NotSurfacePlausibility Thelearnedrankerisasequenceclassifieroverapromptcontainingthecandidate’snodetypeandspan,its text,andupto24numberedcontextlinesoneachside;theinstructionaskswhetherdeletionissafebecause thestatementisunused,andmapsuncertaintyexplicitlytoKEEP.Supervisioncomesonlyfromverifier outcomes yx,c ∈0,1obtainedbyapplyingeachcandidateontheMBPPtrainingrange;testlabelsnever enterprompts,training,orcheckpointselection.Letzθ(x,c)∈Rbetheclassifier’ssuccessmargin,thatis thedifferencebetweenitstwooutputlogits. Foreachtrainingtaskweformagroup Gx ⊆ C(x)containingeveryverifiedpositivePx= c ∈ Gx : yx,c=1 plusselectednegativesupto | Gx |≤ 8,andwe discardtaskswith Px = ∅ ,forwhichno rankingtargetexists.Thelistwiseobjectiveisthenthenegativelogofthesoftmaxmassthatthemodel placesonverifieddeletions, whichisfinitebecause Px ∅andvanishesexactlywhenallmasssitsoncandidatestheverifieraccepts. Equation(6)expressesthedeploymentobjectivemoredirectlythanindependentclassification:ausefulorder needsoneacceptedcandidatenearthefront,notcalibratedprobabilitiesforeveryASTnode.Because(6)is invarianttotherankingamongnegatives,weaddasmallpointwisetermthatstabilizesthemarginscaleand retainsaKEEPsignalonunsuccessfulcandidates, 6 withBCE(z,y)=-ylogσ(z)-(1-y)log(1-σ(z)),σthelogisticfunction,andλ=0.1fixedforall runs.Atdeploymentthesoftmaxsuccessscoreorderscandidates,withthecandidatekeyasadeterministic tie-break;nonatural-languagemodeloutputiseverexecuted. Threecontrolschangethesupportorweightofthesameevidencewhileleaving(7)intact.Exact- residual trainingremovesS3(x)fromGx,askingthemodeltorankonlytheopportunitiesthatcould occupythelearnedtail. Utility trainingrepeatseachverifiedpositive3+min(5, ⌊| c | /80 ⌋ )times,where | c | is itscharacterlength,sothatlargerusefuldeletionsarefavoredbyafactorbetweenthreeandeightwithout lettingonelongspandominateoptimization. Task-marginal trainingisstricterstill:itkeepspositive signalonlyfortaskswithτ(S3,x)=∞.Thatlastobjectivematchesmarginalcoverageconceptuallybut discardsmostpositiveevidencewheneverstaticrankingisstrong,soourexperimentscompareobjectives insteadofassumingthatthemostdeployment-specificlabelistheeasiesttolearn. Thisalsoclarifieswhattheclassifierdoesnotestimate.Itdoesnotpredictsemanticequivalence independentlyofatestsuite,anditdoesnotestimatetheprobabilitythatanarbitraryprojectwilltoleratea deletion.Itsscoreisatask-conditionedrankingstatisticlearnedfromverifieroutcomesinanamedsource domain,andshiftcanchangeboththecandidatemixandtherelationbetweensurfacecontextandtested behavior.Wethereforeuselearnedscorestoorderevidence-producingexperiments,neverasevidence themselves. 4.4ChoosingaSchedule Complementarity,notdominance,iswhatmakesamixtureworthitsslots.ForpoliciesAandB,letUA=x: τ(A,x)<∞.Therelevantquantityisnot|UB|butthedecomposition|UB|-|UA|=|UB |-|UA |: newsuccessesminusdisplacedones.Arankerwithlowerstandalonecoverageremainsvaluablewhenits successesfallontheresidualsetUB\ UA ,andahighstandalonescoredoesnotjustifyreplacementwhen UA .Thisiswhywereportmatcheddiscordancesratherthanonlyaggregatecounts,andwhythe validatedmixture (Equation4)keepsastaticcoreinsteadofhandingallfiveslotstothebetter standaloneranker. Thetwoschedulesthenrepresenttwoexplicitcontracts.Thevalidatedmixtureholdsresourceuse constantandacceptsameasuredriskofdisplacementinexchangeforgreatertarget-domaincoverage;by Proposition2,itsempiricalclaimisconfinedtodomainswithrepresentativevalidationdata.prefix-preserving augmentationacceptsalongerworst-caselist—and,aswequantifybelow,moreverifiercalls—but makespriorcoverageinvarianttothelearnedorderunderadeterministicverifier.Neithercontractmakes adeletionsemanticallysafebeyondthetests.Theirvalueliesinmakingthesourceofuncertaintyexplicit: targetevidencedecideswhetherslotsmaybereplaced,andtheverifierdecideswhetheranyproposededit maybecommitted. 4.5ExecutionProcedureandAuditTrail Foreachtask,theschedulermaterializestheorderedlistbeforeanypatchruns.De-duplicationisstable: whenthelearnedandstaticordersnamethesamespan,theearlierstaticoccurrenceownstheslotandthe learnedlistadvances.Everyevaluationistransactional—copythesource,deletethecandidate’scompleteline span,compile,executethedomainsuiteinanisolatedprocess,anddiscardfailures.Thefirstpassingpatchis committedandlatercandidatesarenevertested.Aprogramthatfailsitsownbaselinesuiteisineligibleand contributesnodeletion. 5ExperimentalProtocol Weaskfourquestions:whetheronestatic-firstarchitectureimprovesfixed-budgetcoverageacrossmodel scalesandrepeatedruns;whetherthegaincomesfromaspeciallearningobjectiveorfromcomplementary 7 Backbone Static Mixture ∆(range) Qwen2.5-0.5B 70.075.7 +5.7(+3,+7) Qwen3-0.6B 70.577.5 +7.0(+6,+9) Qwen3-8B 71.0 78.5 +7.5(+6,+9) Allnineruns 70.477.1 +6.7 Table1:MBPPreplications:meanaccepted-taskcountsoutof499–500eligibleprograms,atamatchedfive-proposal budget.Per-scaledifferencesinparenthesesaretheminimumandmaximumpaireddifferences. scheduling;howlearnedreplacementbehavesundershiftandwhetherprefixpreservationrepairsit;and whichconclusionssurvivestrongerpublicsuites.MBPP(Austinetal.,2021)tasks601–974supplytraining labels,tasks511–600formthedisjointvalidationsliceusedforcheckpointselection,andtasks1–500are theheld-outtest,ofwhich499or500areeligibledependingonbaselineexecutability.Thesame validatedmixture isevaluatedinnineindependentrunsspanningQwen2.5-0.5B,Qwen3-0.6B,andQwen3 -8BrankerswithLoRAadapters.Forshift,MBPP-trained0.5BrankingsareevaluatedonDS-1000(Lai etal.,2023),HumanEval(Chenetal.,2021a),andBigCodeBench-Hard(Zhuoetal.,2025)undertheir officialexecutionoracles;prefix-preservingaugmentationwasdesignedafterthoseresultsandis thereforeexploratorythere,afterwhichitwasfrozenandevaluatedonce,withoutanyadjustment,on HumanEval+(138eligible).MBPP+(170eligible)separatelytestswhetherordinary-MBPPdeletions surviveastrongerin-domainspecification;bothPlussuitesareprovidedbyEvalPlus(Liuetal.,2023). Per-runledgersandsupplementaryanalysesappearinthesupplementarymaterial. 6Results 6.1Fixed-BudgetGainsReplicateAcrossModelScales Table1aggregatestheninematchedruns.Everypairedcomparisonfavorsthemixture,withper-rungains of+3to+9acceptedtasks,thatis4.3%to12.7%relative.Meancoverageimprovesby8.1%,9.9%,and 10.6%relativeforthe0.5B,0.6B,and8Brankers,anall-scalemeanof9.5%(70.4 → 77.1acceptedtasks, or14.1%→15.4%absolutecoverage).Thegainispositiveinallninematchedrunsandconsistentacross modelscales. 6.2WhichProposalsDeservetheSlots Figure2plotseachindividualrankingforK=1...5.Theordersarefarfrominterchangeable:candidate- featurelogisticrankingisstrongestatK=1(43acceptedversus31forshortest),listwiserankingisweakest there(27)yetovertakesshortestbyK=2,andutilityrankingremovesthemostcharactersateveryK≥2. Wekeepthispanelseparatefromthedeployedcomparisonbecausetheexperimentrecordedthepreregistered K=5mixtureratherthanaretrospectivelycompletedmixturecurve. Atthematchedbudgetoffiveproposals,Table2showsthatallthreemixturesimprovecoverageand characterreductionoverthefive-candidateshortest-firstbaseline,anddosowithoutspendingmoreverification: thecandidate-featuretailraisesaccepteddeletionsby14.3%using12fewerverifiercalls,andtheutility-ranked tailreachesthestrongestsame-protocolpointat18.6%relativegain.Verifieddeletionsperthousandverifier callsrisefrom43.9to52.5,a19.5%efficiencyimprovement,becauseaproposalthatsucceedsearlyalso stopsthescheduleearly.Theutilitypointmarkstheattainablefrontier;itdoesnotestablishutilityweighting asuniversallypreferable. 8 Policy(|π|=5) Accept CallsAcc./1kChar.red. Staticshortest-5701,59343.9 2.11% Mixture,candidateLR801,58150.6 2.84% Mixture,listwise781,58649.2 2.99% Mixture,utility 83 1,581 52. 5 3.06% Table2:MatchedMBPPcomparisonatacommonfive-proposalbudgetover500tasks.Eachmixturetestsathree- candidatestaticcorefollowedbytwolearnedcandidates.“Calls”istotalverifierinvocationsand“Acc./1k”accepted deletionsperthousandcalls,sothemixturesimprovecoveragewithoutbuyingitwithextraverification. Matcheddiscordancesmakecomplementarityconcrete.Theutility-rankedmixturehas14successes absentfromthestaticbaselinewhilethebaselinehasoneabsentfromit.Againstthecandidate-feature mixture,theutilityvariantaddsthreetasksandlosesnone.Thefrontierisreal;theidentityofthebest objectiveisnotresolvedbythesedata. 6.3TheGainIsArchitectural,NotanObjectiveMiracle On0.5Bvalidation,globallistwiseandexact-residualtrainingtieatameanbestscoreof8.75acceptedtasks, andonthefrozentestseedsavailableforboth,globallistwiseaverages2.6%moreaccepteddeletionsthan exactresidual.Thetask-marginalobjectivepeakedat9acceptedvalidationtasksinallthreeofitsseeds againstapromotiongateof10—theleveltheadvancedarmsreached—andwasthereforeneverevaluated ontest,adecisionmadebeforeanytestaccess.Itssupervisionissparsebyconstruction,sinceonlytasks withaverifiedsuccessoutsidethethree-candidatestaticcorecontributepositivemarginalsignal.Together thesecontrolsruleoutresidual-onlytrainingasthesourceofthegainandleavethearchitecturalexplanation: staticproposalscovereasycandidates,learnedproposalscontributeadifferentorder,andexecutionfiltersthe incorrectones. Candidatecompositionshowsthereallocationdirectly.AtK=5theshortestorderaccepts35 assignmentsbutonly19conditionals,withmeandeletionlength1.39lines.Candidate-featurerankingaccepts 33conditionalsand30assignmentsat1.62lines,andlistwiserankingaccepts32conditionalswhileremoving 3.16%ofcorpuscharacters.Thecandidate-featuremixturekeepsall35assignmentsuccessesvisibletoits staticprefixwhileraisingconditionalsuccessesto31andmeandeletionlengthto1.58.Themodelisnot reproducing“deletetheshortestline”withnoise;itspendsitstwoslotsoncontextualcontrol-flowcandidates, whichispreciselytheregionsoundstaticanalysiscannotcertify—andalsotheregionwithmorewaysto bewrong,whichiswhyexecutionremainsmandatory.Characterreductionconfirmsthatthegainisnot concentratedintrivia:everymixtureimprovesonbothaxesofTable2simultaneously. 6.4UnderShift,PreservethePrefix Model-onlytransferisunreliable(Table3).OnDS-1000learnedcoveragespans9.35–9.70%againsta9.58% zero-referencebaseline;onHumanEvalitswingsfrom59.15%to96.95%arounda93.29%baseline;on BigCodeBench-Harditreachesonly41.22–50.00%against60.14%forshortest-first.Thelearnedordersdo removemorecharacterswhentheysucceed—2.61–3.07%versus1.24%onBigCodeBench-Hard—sothe failureisacoverage–reductiontrade-offratherthanuniformincompetence.Eitherway,arankerthathelps in-domaincannotjustifyreplacingacompletestaticprefixunderunknownshift,exactlyasProposition2 predicts. Prefix-preservingaugmentationchangestheoperationalconclusionwithoutclaimingthescoresbecame calibrated.Learnedreplacementcanpushcoveragebelowthestaticreference,whereasaugmentationwas non-degradinginallninefrozenreplaysandstrictlypositiveineight,improvingmeancoverageby0.23,2.24, 9 ● ShortestCand.LRListwise 令 Utility 80 70 60 50 40 30 20 12345 ProposalbudgetK Figure2:MBPPbudgetcurvesfortheindividualorders,notthemixtures.TherankingsdisagreemostatsmallK, whichiswhatafive-slotschedulehastoarbitrate. and3.38percentagepointsonthethreebenchmarks.Thisisnotfree.Becausetheextraslotsareonlyreached afterfivestaticfailures,theycost62.5%,4.8%,and24.8%moreverifiercallsrespectively—thepriceofthe guarantee,andonDS-1000anunattractiveonefora0.23-pointgain.Thecostisexplicitandquantifiable, andshouldbeweighedagainstthecoverageimprovementinanydeploymentdecision. 6.5StrongerTestsDefinetheSafetyBoundary MBPP+givesadeliberatelyadverseresult.Amongthe170programsthatpassthestrongersuitebefore deletion,thefive-candidateshortest-firstbaselineandallthreestatic-firstmixturesacceptexactly11deletions, thatis6.47%coverage,andthemixturesraisebase-onlyrejections—patchesacceptedbytheoriginalMBPP testsbutrejectedbyMBPP+—from3to7.Theordinary-MBPPadvantagedoesnotsurvivethestronger specification,andthemixtures’additionalproposalsfaildisproportionatelyunderit.“Verified”alwaysmeans verifiedbythestatedoracle. HumanEval+evaluatesthefrozen prefix-preservingaugmentation withouttarget-specificadjustment. Thecompletefive-candidatestaticrankingacceptsdeletionsfor92.75%ofthe138eligibleprograms;across threerunsaugmentedcoverageis97.83%,92.75%,and94.93%,sotworunsimproveandoneties,at4.3 –5.7%extraverifiercalls(Table4).Thereplayreusesfrozenstandard-verifierledgers,whichdonot exposeeveryunsuccessfulsecondarycandidate,sotheaugmentedfiguresareconservativelower boundsratherthanacompletetwo-candidatereplay.Theseresultstesttheorderingpropertyof Proposition1underastrongeroracleandreinforcethat“verified”alwaysmeansverifiedbythestated specification. 10 DatasetElig.StaticLearnedonlyAugm.Calls DS-10008669.589.35–9.709.82 +62.5% HumanEval16493.2959.15–96.9595.53 +4.8% BCB-Hard14860.1441.22–50.0063.51 +24.8% Table3:Accepted-deletioncoverage(%)underdistributionshiftforthelocallystrongeststaticorder(zero-referenceon DS-1000andHumanEval,shortest-firstonBigCodeBench-Hard),thethreelearned-onlyrankings,andprefix- preservingaugmentation(meanoverthesamethreeruns).Thelastcolumnistheaugmentedpolicy’sverifier-call overheadrelativetothestaticorder.Thisshifted-benchmarkanalysisisexploratory. Verifier PolicyCoverage Char.red. MBPP+ staticshortest-56.47%0.76% MBPP+ mixtures(allthree) 6.47%0.78% HE+ completestatic-5 92.75%56.28% HE+augmented,3runs92.75–97.83%56.28–61.21% Table4:Strong-verifierboundaryon170MBPP+and138HumanEval+(HE+)eligibleprograms.TheMBPP+tie holdsforallthreemixtures,whichalsoraisebase-onlyrejectionsfrom3to7. 7Discussion AIcodinghasfundamentallyalteredtheeconomicsofsoftwarecreation.Whenfrontiermodelsproduce benchmark-correctsolutionsatorabovestronghumanreferencepoints,generationceasestobethescarce operation;understanding,consolidation,andremovalbecomecomparativelymorevaluable.Vibecoding sharpensthisimbalancebecauseeachpromptaddsalocalsolutionwhilepreservingearlierscaffolding.The resultingredundancyisnotmerelyaesthetic:itexpandsreviewsurface,multipliesthestatesfutureagents mustreasonabout,andconvertsshort-termgenerationspeedintolong-termdebt.Deletiondeservestobea first-classAIcodingcapabilityratherthananoccasionalmanualcleanup. Thedesignchoicewedefendmoststronglyisseparatingdiscoveryfromauthority,andthenbeingexplicit aboutwhichevidencelicenseswhichschedule.Learnedrankersrecognizecontextualredundancythatstatic ordersmiss,buttheirscoresareproposals,notcorrectnessevidence.Withrepresentativevalidationdataand afixedfive-slotbudget,thevalidatedmixturetradestwostaticslotsforcontextualsearch;withoutsuch data, prefix-preservingaugmentation teststhefullstaticrankingfirstandconsultslearnedcandidatesonly afterallfivefail,sothelearnedtailcandiscoverextraremovalsbutcanneversuppressadeletionthe establishedpolicywouldhavefound.Whatdecidesbetweenthemisavailabletargetevidenceandtherequired guarantee—notavaguejudgmentabouthowmuchtotrustthemodel. Theabstractionextendspaststatementdeletion.RepositoryagentsandIDEscoulduseittoretire duplicatedhelpersaftergeneration,simplifypatchesbeforereview,removecompatibilityscaffoldingafter migrations,orprioritizerefactoringcandidatesincontinuousintegration:aconservativeanalyzeroccupies aprotectedprefix,alearnedmodelexplorescomplementarycandidates,andtheproject’sowntestsgate commitment.Largermodelsmayimprovethelearnedtail,buttheycannotsubstitutefortestadequacy,and ourMBPP+resultofferstheclearestdemonstrationofthatlimit.Thepracticalopportunityisamaintenance loopinwhichAIsystemsnotonlygeneratesoftwarebutcontinuouslyjustifyandremovewhattheprojectno longerneeds. 11 8Limitation Ourlabelsandtargetsarebenchmarkprogramsratherthanmaintenancehistories,andtheeffectsizes aremodestinabsoluteterms:coveragemovesfromabout14%to15%ofMBPPprograms.Flakyor environment-dependenttestsbreakthedeterminismthatProposition1assumes,soadeploymentshould pindependencies,isolateexecution,andrepeatnondeterministictests.Finally,deletingcodecanremove logging,monitoring,securitychecks,orrarebehaviorthatnotestexercises.Humanreviewandadequate specificationsremainnecessary;DelScoutmustnotbepresentedasasemanticproof,norusedtoweaken safety-criticalcodemerelybecauseabenchmark-styleverifierpassed. 9Conclusion ThenextfrontierofAIcodingisnotonlygeneratingmorecodebutkeepinggeneratedsoftwarecomprehensible asitevolves.Weformulatedthatmaintenanceproblemasverifier-backeddeletionunderafiniteproposal budgetandintroducedDelScout,whichcombinesdeterministicstaticcandidateswithcontext-sensitive learnedcandidateswhilereservingacceptanceforexecutiontests.Asharedstatic-firstmixtureimproves MBPPcoverageacrossthreemodelscalesandninerunswithoutextraverificationcost,andobjective controlsattributethebenefittocomplementaryschedulingratherthanaprivilegedtrainingloss.Undershift, replacingstaticcandidatescanfail,andretainingthecompletestaticprefixrestoresnon-decreasingcoverage byconstruction—ataverifier-callcostwemeasureratherthanhide.StrongerPlussuitesthenmarkthe boundary:schedulinggovernssearch,butonlythespecificationgovernswhat“preservingbehavior”means. Thegoverningprincipleisdeliberatelysimple.Letmodelsbroadenthesearchforunnecessarycode,let orderconstraintheirfailuremodes,andletprojectevidencedecidewhatmaybedeleted.ReliableAIcoding isnotcompletewhenasystemcanwriteaworkingprogram;itiscompleteonlywhenthesystemcanalso helpthatprogramstaynolargerthanitneedstobe. 12 Appendix AExperimentalConfiguration MBPP(Austinetal.,2021)tasks601–974providetraininglabels,tasks511–600formthedisjointvalidation sliceusedforcheckpointselection,andtasks1–500formtheheld-outtest;499or500programsareeligible dependingonbaselineexecutability.Thesamestatic-firstmixture—threedeterministicshortest-firstcandidates followedbytwolearnedcandidatesabsentfromthatcore,withinafive-proposalbudget—isevaluatedinnine independentrunsacrossQwen2.5-0.5B,Qwen3-0.6B,andQwen3-8Brankers.Distribution-shiftevaluation usesDS-1000(Laietal.,2023),HumanEval(Chenetal.,2021a),andBigCodeBench-Hard(Zhuoetal., 2025)withtheirofficialexecutionsuites;MBPP+andHumanEval+fromEvalPlus(Liuetal.,2023)provide strongerpublicspecifications. EachrankerisasequenceclassifierwithtwooutputheadsandaLoRAadapterofrank8andscaling16, dropout0.05,appliedtothequeryandvalueprojections,withtheclassificationheadalsotrainable.Training usesAdamWatlearningrate2 × 10-5,gradientclippingatnorm1.0,gradientaccumulationoverfourtask groups,amaximumpromptlengthof768tokens,andfourepochs,ofwhichvalidationselectsone.Task groupscontaineveryverifiedpositiveplusnegativesuptoagroupsizeofeight;negativesareorderedto preferthenodetypesthatappearamongthattask’spositives,thenshorterspans,sothelistwisecontrastisnot dominatedbytriviallydissimilarcandidates.Thepromptcontainsthecandidate’snodetypeandspan,its text,andupto24numberedcontextlinesbeforeandafter,andtheinstructionexplicitlymapsuncertaintyto KEEP.Atdeploymentthesoftmaxsuccessscoreorderscandidateswiththecandidatekeyasadeterministic tie-break;nonatural-languagemodeloutputisexecuted. Selected8Bcheckpointswererescoredtwiceoverall4,863candidatesandtheresultingcandidateorders werebitwiseidentical(SHA-256recordedperseed).Validationselections,task-leveloutcomes,andanalysis inputsareretainedinmachine-readableform,soeveryreportedaggregatecanberegeneratedwithoutexposing environment-specificpathsoridentifiers. BFormalInterpretationoftheScheduler LetDbetheeligibleprogramset.Foreachx∈Dapolicyreturnsanordered,duplicate-freelist π(x)=(c1,...,cm);candidatecjdeletesonecompletestatement,producingx \ cj,andthedeterministic verifierVreturns1exactlywhenthatpatchedprogramcompilesandpassesthenamedsuite.Withmin∅=∞, definethepassing-indexsetandtheacceptedindex JV (π,x)=j∈1,...,m:V(x )=1,(8) τ V (π,x)=minJV(π,x),(9) soτ V (π,x)=∞meansthepolicycommitsnodeletion.Verified-deletioncoverageis Letℓ(x)bethenumberofsourcecharactersandr(x,c)=(ℓ(x) - ℓ(x \ c))/ℓ(x).Per-taskcharacterreduction isdefinedwithoutreferringtoanundefinedc∞: 13 andthedataset-levelreductionisRed V x RV (π,x).Afailedtaskthereforecontributeszero andasuccessfultaskcontributesonlyitsfirstverifier-passingdeletion. WriteSk(x)=(s1,...,sk)forthefirstkentriesofthestaticorderandM A (x)forthefirstBentries ofthelearnedorderthatdonotbelongtothecandidatesetA; ∥ denotesorderedconcatenation. Thetarget-validatedfixed-budgetpolicyandtheshiftpolicyare πmix(x)=S3(x)∥M S3(x)(x) , |πmix| =5,(12) πaug(x)=S5(x) ∥ M S5(x)(x) , | πaug | ≤ 5+B.(13) De-duplicationistakenagainstthestaticcandidateseachpolicyactuallyproposes: S3 forthemixture,whose staticslots4and5aretheonesbeingexchanged,andS5foraugmentation.Thefirstpolicyreplacestwostatic slotsonlywhentarget-domainvalidationsupportstheexchange;thesecondneverreplacesastaticproposal andonlyaddssearchbeyondthecompleteprefix. Proofofprefixpreservation.Thefirstfiveentriesofπaug(x)areexactlyS5(x).Ifτ V (S5,x)=j ≤5,thenevaluationofπ aug teststheidenticalcandidatess1,...,sjintheidenticalorder,allofs1,...,sj -1failasbeforebecauseVisdeterministic,andsjiscommitted;henceτ V (πaug,x)=jwiththesame acceptedcandidate.Ifτ V (S5,x)=∞,augmentationeitheralsofailsorsucceedssomewhereinits learnedtail.Inbothcases 1[τ V (πaug,x)< ∞ ] ≥ 1[τ V (S5,x)< ∞ ],(14) andaveragingthepointwiseinequalitygivesCovV(πaug)≥CovV(S5).Becausestaticsuccesses retainthesameacceptedcandidateandeverynewlearnedsuccesshasr(x,c) ≥ 0,thesame argumentgivesRedV (πaug) ≥RedV(S5).Bothstatementsrequiredeterministicexecutionand constrainonlybehaviorrepresentedbyV.Theysaynothingaboutthenumberofverifiercalls,whichstrictly increasesontaskswherethestaticprefixfails;thatcostisreportedinthemainpaper. Whyreplacementneedsvalidation.Supposeafixed-budgetmixtureomitssome sj ∈ SK . ConsideraninstanceonwhichV(x )=1whileeveryretainedstaticcandidateandeverylearned candidatefails.ThenSKsucceedsandthemixturedoesnot,sononontrivialreplacementpolicy dominatesthecompletestaticprefixoneverytargetdistribution.Whatgovernstheobservedgainis insteadcomplementarity:with Uπ =x:τ V (π,x)<∞, |Uπ|-|US|=|Uπ |-|US π|.(15) Thefirsttermcountsnewsuccessesandthesecondcountsdisplacedstaticsuccesses.Prefix-preserving augmentationforcesthesecondtermtozerobyconstruction,whereasafixed-budgetmixturemustestimate bothtermsonrepresentativevalidationdata. CCompleteMBPPReplications Theper-rundetailmattersforinterpretation.Thedirectionalresultisconsistentacrossallnineruns.Shortest- firstitselfaccepts70or71tasksdependingonwhichfrozenrunisread,becauseoneprogram’seligibility dependsonbaselineexecutability;allcomparisonsarethereforepairedwithinarun. DDistributionShift,PerSeed Twopropertiesofthisreplaybounditsinterpretation.First,theaugmentedarmisreconstructedfromfrozen official-verifierledgers:alearnedcandidateiscreditedonlyifthelearned-onlyrunhadalreadyacceptedit 14 Backbone ObjectiveStatic Mixture ∆ Qwen2.5-0.5B residual70 77 +7 Qwen2.5-0.5B residual 7073 +3 Qwen2.5-0.5B residual 7077 +7 Qwen3-0.6B listwise71 77 +6 Qwen3-0.6B listwise 7180 +9 Qwen3-0.6B listwise7076 +6 Qwen3-0.6B listwise 7077 +7 Qwen3-8B residual 71 80 +9 Qwen3-8B residual7177 +6 Allnine — 70.4 77.1 +6.7 Table5:Everyunlockedreplication,inaccepted-taskcountsontheheld-outMBPPtest.Allninepaireddifferencesare positive. DatasetStatic LearnedonlyAugmented DS-1000(866) 83 81/82/84 85/84/86 HumanEval(164) 153159/97/148161/153/156 BCB-Hard(148) 89 61/68/74 93/95/94 Table6:Accepteddeletionsperseedundershift,witheligibleprogramsinparentheses.Staticisthelocallystrongest staticorder(zero-referenceonDS-1000andHumanEval,shortest-firstonBigCodeBench-Hard).Verifiercallsrise from1,951to3,171onDS-1000,from413to433onHumanEval,andfrom468to584onBigCodeBench-Hard,i.e. +62.5%,+4.8%,and+24.8%onaverage. withinitsownfirstBattempts,soacceptedcountsarelowerbounds.Second,thereplaychargesuptoB learnedattemptswithoutremovingoverlapswiththealready-failedstaticprefix.Suchanoverlapcannotcreate anaccepteddeletion—thesamecandidateunderthesamedeterministicverifierfailsagain—sotheguarantee isunaffected,andthereportedcalloverheadsareupperboundsonwhatade-duplicatedimplementation wouldspend. Theguaranteealsoassumesdeterministicexecution.Underaflakyverifierthesamestaticcandidatecan changeoutcomebetweenschedulesevenwhenitspositionispreserved,soapracticalimplementationshould pindependencies,isolateexecution,andrepeatnondeterministictests. EStrongerVerifiers OntheMBPP+intersection(170programspassingthestrongersuitebeforedeletion)shortest-firstand allthreestatic-firstmixturesacceptexactly11deletions.Themixturesrecordsevenbase-onlyrejections againstthreeforshortest-first,andthelearned-onlyordersaccept10or11withsixbase-onlyrejections.The ordinary-MBPPadvantagethereforedoesnotsurvivethestrongerspecification. HumanEval+contributesthe138of164programsthatpassboththebaseandPlussuitesbeforedeletion. Thecompletefive-candidatezero-referencerankingisevaluatedfirst,andatmosttwolearnedproposalsare consideredonlyafterallfivefail.Staticcoverageis128/138(92.75%);thethreeaugmentedrunsobtain135, 128,and131accepteddeletions(97.83%,92.75%,94.93%)withcharacterreduction61.21%,56.28%,and 58.88%against56.28%,at363,368,and367verifiercallsagainst348. 15 References JacobAustin,AugustusOdena,MaxwellNye,MaartenBosma,HenrykMichalewski,DavidDohan,Ellen Jiang,CarrieCai,MichaelTerry,QuocLe,andCharlesSutton.Programsynthesiswithlargelanguage models. arXivpreprintarXiv:2108.07732 ,2021. BeiChen,FengjiZhang,AnhNguyen,DaoguangZan,ZeqiLin,Jian-GuangLou,andWeizhuChen.CodeT: Codegenerationwithgeneratedtests.InInternationalConferenceonLearningRepresentations, 2023. MarkChen,JerryTworek,HeewooJun,QimingYuan,HenriquePondedeOliveiraPinto,JaredKaplan,Harri Edwards,YuriBurda,NicholasJoseph,GregBrockman,etal.Evaluatinglargelanguagemodelstrainedon code.arXivpreprintarXiv:2107.03374,2021a. XinyunChen,MaxwellLin,NathanaelSchärli,andDennyZhou.Teachinglargelanguagemodelsto self-debug.InInternationalConferenceonLearningRepresentations,2024. ZiminChen,SteveKommrusch,MicheleTufano,Louis-NoëlPouchet,DenysPoshyvanyk,andMartin Monperrus.SequenceR:Sequence-to-sequencelearningforend-to-endprogramrepair.IEEE Transactions onSoftwareEngineering,47(9):1943–1959,2021b. DeepSeek-AI.DeepSeek-R1:IncentivizingreasoningcapabilityinLLMsviareinforcementlearning. arXivpreprintarXiv:2501.12948,2025. SebastianEder,MaximilianJunker,ElmarJürgens,BenediktHauptmann,RudolfVaas,andKarl-Heinz Prommer.Howmuchdoesunusedcodematterformaintenance?InProceedingsofthe34th InternationalConferenceonSoftwareEngineering ,pages1102–1111,2012. ZhangyinFeng,DayaGuo,DuyuTang,NanDuan,XiaochengFeng,MingGong,LinjunShou,BingQin, TingLiu,DaxinJiang,andMingZhou.CodeBERT:Apre-trainedmodelforprogrammingandnatural languages.InFindingsoftheAssociationforComputationalLinguistics:EMNLP,pages1536–1547, 2020. MartinFowler. Refactoring:ImprovingtheDesignofExistingCode .Addison-Wesley,1999. DanielFried,ArmenAghajanyan,JessyLin,SidaWang,EricWallace,FredaShi,RuiqiZhong,Wentau Yih,LukeZettlemoyer,andMikeLewis.InCoder:Agenerativemodelforcodeinfillingandsynthesis.In InternationalConferenceonLearningRepresentations,2023. ClaireLeGoues,MichaelDewey-Vogt,StephanieForrest,andWestleyWeimer.Asystematicstudyof automatedprogramrepair:Fixing55outof105bugsforUS$8each.InProceedingsofthe 34thInternationalConferenceonSoftwareEngineering,pages3–13,2012. DayaGuo,ShuoRen,ShuaiLu,ZhangyinFeng,DuyuTang,ShujieLiu,LongZhou,NanDuan,Alexey Svyatkovskiy,ShengyuFu,MicheleTufano,ShaoKunDeng,ColinClement,DawnDrain,NeelSundaresan, JianYin,DaxinJiang,andMingZhou.GraphCodeBERT:Pre-trainingcoderepresentationswithdataflow. InInternationalConferenceonLearningRepresentations,2021. EdwardJ.Hu,YelongShen,PhillipWallis,ZeyuanAllen-Zhu,YuanzhiLi,SheanWang,LuWang,and WeizhuChen.LoRA:Low-rankadaptationoflargelanguagemodels.In InternationalConference onLearningRepresentations,2022. BinyuanHui,JianYang,ZeyuCui,JiaxiYang,DayihengLiu,LeiZhang,TianyuLiu,JiajunZhang,Bowen Yu,KemingLu,etal.Qwen2.5-Codertechnicalreport.arXivpreprintarXiv:2409.12186,2024. 16 CarlosE.Jimenez,JohnYang,AlexanderWettig,ShunyuYao,KexinPei,OfirPress,andKarthikNarasimhan. SWE-bench:Canlanguagemodelsresolvereal-worldGitHubissues?InInternationalConference onLearningRepresentations,2024. JensKnoop,OliverRüthing,andBernhardSteffen.Partialdeadcodeelimination.InProceedingsofthe ACM SIGPLANConferenceonProgrammingLanguageDesignandImplementation,pages147–158, 1994. PhilippeKruchten,RobertL.Nord,andIpekOzkaya.Technicaldebt:Frommetaphortotheoryandpractice. IEEESoftware,29(6):18–21,2012. YuhangLai,ChengxiLi,YimingWang,TianyiZhang,RuiqiZhong,LukeZettlemoyer,WentauYih,Daniel Fried,SidaWang,andTaoYu.DS-1000:Anaturalandreliablebenchmarkfordatasciencecodegeneration. InProceedingsofthe40thInternationalConferenceonMachineLearning,pages18319–18345, 2023. HungLe,YueWang,AkhileshDeepakGotmare,SilvioSavarese,andStevenC.H.Hoi.CodeRL:Mastering codegenerationthroughpretrainedmodelsanddeepreinforcementlearning.In Advancesin NeuralInformationProcessingSystems,volume35,pages21314–21328,2022. RaymondLi,LoubnaBenAllal,YangtianZi,NiklasMuennighoff,DenisKocetkov,ChenghaoMou,Marc Marone,ChristopherAkiki,JiaLi,JennyChim,etal.StarCoder:Maythesourcebewithyou! TransactionsonMachineLearningResearch ,2023. YujiaLi,DavidChoi,JunyoungChung,NateKushman,JulianSchrittwieser,RémiLeblond,TomEccles, JamesKeeling,FelixGimeno,AgustinDalLago,etal.Competition-levelcodegenerationwithAlphaCode. Science,378(6624):1092–1097,2022. JiaweiLiu,ChunqiuStevenXia,YuyaoWang,andLingmingZhang.IsyourcodegeneratedbyChatGPT reallycorrect?rigorousevaluationoflargelanguagemodelsforcodegeneration.In Advancesin NeuralInformationProcessingSystems,volume36,2023. KuiLiu,AnilKoyuncu,DongsunKim,TegawendéF.Bissyandé,TiagoDurieux,LizhenWu,JacquesKlein, andYvesLeTraon.TBar:Revisitingtemplate-basedautomatedprogramrepair.In Proceedingsof the28thACMSIGSOFTInternationalSymposiumonSoftwareTestingandAnalysis,pages31–42, 2019. FanLongandMartinRinard.Automaticpatchgenerationbylearningcorrectcode.InProceedingsofthe 43rdAnnualACMSIGPLAN-SIGACTSymposiumonPrinciplesofProgrammingLanguages, pages298–312,2016. StephanLukasczyk,FlorianKroiß,andGordonFraser.Pynguin:Automatedunittestgenerationforpython. In Proceedingsofthe44thInternationalConferenceonSoftwareEngineering:Companion Proceedings,pages168–172,2022. SergeyMechtaev,JooyongYi,andAbhikRoychoudhury.Angelix:Scalablemultilineprogrampatchsynthesis viasymbolicanalysis.InProceedingsofthe38thInternationalConferenceonSoftwareEngineering, pages691–701,2016. GhassanMisherghiandZhendongSu.HDD:Hierarchicaldeltadebugging.In Proceedingsofthe IEEE/ACMInternationalConferenceonAutomatedSoftwareEngineering ,pages142–151,2006. AnsongNi,SriniIyer,DragomirRadev,VesStoyanov,WentauYih,SidaWang,andXiVictoriaLin.LEVER: Learningtoverifylanguage-to-codegenerationwithexecution.In Proceedingsofthe40th InternationalConferenceonMachineLearning ,pages26106–26128,2023. 17 JohnRegehr,YangChen,PascalCuoq,EricEide,ChuckyEllison,andXuejunYang.Test-casereductionfor Ccompilerbugs.In ProceedingsoftheACMSIGPLANConferenceonProgrammingLanguage DesignandImplementation,pages335–346,2012. SimoneRomano,ChristopherVendome,GiuseppeScanniello,andDenysPoshyvanyk.Amulti-study investigationintodeadcode.IEEETransactionsonSoftwareEngineering,46(1):71–99,2020. AndréSilva,SenFang,andMartinMonperrus.RepairLLaMA:Efficientrepresentationsandfine-tuned adaptersforprogramrepair.InProceedingsofthe46thInternationalConferenceonSoftware Engineering ,2024. ChengnianSun,YuanboLi,QirunZhang,TianxiaoGu,andZhendongSu.Perses:Syntax-guidedprogram reduction.InProceedingsofthe40thInternationalConferenceonSoftwareEngineering,pages 361–371,2018. YueWang,WeishiWang,ShafiqJoty,andStevenC.H.Hoi.CodeT5:Identifier-awareunifiedpre-trained encoder-decodermodelsforcodeunderstandingandgeneration.In Proceedingsofthe2021 ConferenceonEmpiricalMethodsinNaturalLanguageProcessing,pages8696–8708,2021. YueWang,HungLe,AkhileshDeepakGotmare,NghiD.Q.Bui,JunnanLi,andStevenC.H.Hoi.CodeT5+: Opencodelargelanguagemodelsforcodeunderstandingandgeneration.InProceedingsofthe 2023ConferenceonEmpiricalMethodsinNaturalLanguageProcessing,pages1069–1088,2023. WestleyWeimer,ThanhVuNguyen,ClaireLeGoues,andStephanieForrest.Automaticallyfindingpatches usinggeneticprogramming.InProceedingsofthe31stInternationalConferenceonSoftware Engineering,pages364–374,2009. ChunqiuStevenXiaandLingmingZhang.Lesstraining,morerepairingplease:Revisitingautomated programrepairviazero-shotlearning.InProceedingsofthe37thIEEE/ACMInternational ConferenceonAutomatedSoftwareEngineering ,pages1–13,2022. AndreasZellerandRalfHildebrandt.Simplifyingandisolatingfailure-inducinginput.IEEETransactions onSoftwareEngineering,28(2):183–200,2002. QihaoZhu,ZeyuSun,YuananXiao,WenjieZhang,KangYuan,YingfeiXiong,andLuZhang.A syntax-guidededitdecoderforneuralprogramrepair.InProceedingsofthe29thACMJoint European SoftwareEngineeringConferenceandSymposiumontheFoundationsofSoftware Engineering,pages341–353,2021. TerryYueZhuo,MinhChienVu,JennyChim,HanHu,WenhaoYu,RatnadiraWidyasari,ImamNurBani Yusuf,HaolanZhan,JundaHe,IndraneilPaul,etal.BigCodeBench:Benchmarkingcodegenerationwith diversefunctioncallsandcomplexinstructions.InInternationalConferenceonLearning Representations,2025.

Instant research discovery