CounterBench: Evaluating and Improving Counterfactual Reasoning in | | | Large | Language | Models | | | | | | | | --- | ------------------ | ------------- | ----------------------------- | ------- | --- | -------------- | ----- | --- | --- | --- | | | YuefeiChen1 | VivekK.Singh1 | | JingMa2 | | RuixiangTang1* | | | | | | | 1RutgersUniversity | | 2CaseWesternReserveUniversity | | | | | | | | | | Abstract | | | | | | Back. | | | | 6202 rpA 11 ]LC.sc[ 2v80011.2052:viXra Counterfactualreasoningiswidelyrecognized 74.5 | | | | | | Basic | | 58.5 | | Nested | | | ------------------------------------------- | -------------------- | ------------- | --- | --- | ----- | ------------- | ----- | -------- | ------ | --- | | as one of | the most challenging | and intricate | | | | | | | | | | | | | | | | 98.4 75.272.4 | 53. 0 | 76.078.4 | 94.0 | | | aspectsofcausalityinartificialintelligence. | | | In | | | | | | | | 54.8 5 0.8 thispaper,weevaluatetheperformanceoflarge 5 8 .8 65 . 6 50.8 | languagemodels(LLMs)incounterfactualrea- | | | | | | | | 67.2 | | | | ---------------------------------------- | --- | --- | --- | --- | --- | --- | ---- | ---- | --- | --- | | | | | | | | | 70.8 | 72.4 | | | soning. Incontrasttopreviousstudiesthatpri- | | | | | | | 89.6 | | 90.0 | | | | ------------------------------------- | --- | --- | --- | --- | --- | ----- | --- | ----- | --- | --- | | marilyfocusoncommonsensecausalreason- | | | | | | Cond. | | Joint | | | ing,whereLLMsoftenrelyonpriorknowledge | | | | | | | Our CoIn | | Standard | | | | --- | --- | --- | --- | --- | --- | -------- | --- | -------- | --- | --- | forinference,wespecificallyassesstheirabil- | | | | | | | CausalCoT | | Solver | | | | --- | --- | --- | --- | --- | --- | --------- | --- | ------ | --- | --- | itytoperformcounterfactualinferenceusinga | setofformalrules. | Tosupportthisevaluation, | | | | | | | | | | | ----------------- | ------------------------ | --- | --- | --- | --- | --- | --- | --- | --- | --- | Figure1: ComparisonofaccuracyscoresontheCoun- weintroduceanewbenchmarkdataset,Coun- terBench dataset across different strategies: our pro- terBench,comprising1.2Kcounterfactualrea- posedCoInparadigmversusbaselineapproaches(Stan- | soningquestions. | Thedatasetisdesignedwith | | | | | | | | | | | ---------------- | ------------------------ | --- | --- | ----- | --------- | ---- | ------- | ------ | ---------- | ---- | | | | | | dard, | CausalCoT | (Jin | et al., | 2023), | and Solver | (Hua | varyinglevelsofdifficulty,diversecausalgraph et al., 2024)), evaluated using Gemini-1.5-flash. Our structures,distincttypesofcounterfactualques- | | | | | CounterBenchdatasetincludesfivekindstypes. | | | | | | Basic | | --- | --- | --- | --- | ------------------------------------------ | --- | --- | --- | --- | --- | ----- | tions,andmultiplenonsensicalnamevariants. focusesonexploringhowasinglechangeinacausal Ourexperimentsdemonstratethatcounterfac- variable. Jointinvolvessimultaneouschangesinmul- tualreasoningposesasignificantchallengefor tiplecauses,Nestedinvolvesstepwisehypotheticalas- LLMs,withmostmodelsperformingatlevels sumptionsaboutmultiplevariables. Conditionaleval- | comparabletorandomguessing. | | Toenhance | | | | | | | | | | --------------------------- | --- | --------- | --- | -------------------------------------------- | --- | --- | --- | --- | --- | --- | | | | | | uatescounterfactualsunderobservedconditions. | | | | | | And | LLM’scounterfactualreasoningability,wepro- Backdoorinvolvescounterfactualreasoninginthepres- poseanovelreasoningparadigm,CoIn,which enceofbackdoorpathsthatcreateconfoundingbetween guidesLLMsthroughiterativereasoningand thetreatmentvariableandtheoutcome. backtrackingtosystematicallyexplorecounter- | factual solutions. | Experimental | results | show | | | | | | | | | ------------------ | -------------------- | -------- | ---- | ------ | ------- | ----- | --- | ------- | ----- | ------- | | that our | method significantly | improves | LLM | | | | | | | | | | | | | Koonce | et al., | 2011; | Gow | et al., | 2016; | Loi and | performanceoncounterfactualreasoningtasks | | | | | Rodrigues,2012). | | | Forexample,aconsumerwho | | | | | --- | --- | --- | --- | ---------------- | --- | --- | ----------------------- | --- | --- | --- | andconsistentlyenhancesperformanceacross declinedanextendedwarrantymaylaterwonder, | different | LLMs. Our dataset | is available | at | | | | | | | | | --------- | ----------------- | ------------ | --- | --- | --- | --- | --- | --- | --- | --- | “WhatifIhadpurchasedit,couldIhaveavoidedthe https://huggingface.co/datasets/CounterBench /CounterBench. repaircosts?” Thisillustrateshowcounterfactual | | | | | reasoning | guides | decision-making | | | by | evaluating | | -------------- | --- | --- | --- | --------------------------------------------- | ------ | --------------- | --- | ------ | ----- | ---------- | | 1 Introduction | | | | missedopportunitiesandalternativeoutcomes(Kr- | | | | | | | | | | | | ishnamurthy | | and Sivaraman, | | 2002). | While | tradi- | Counterfactualreasoning,residingatthepinnacle tionalcausalinferencemethods(SharmaandKici- ofPearl’sCausalHierarchy(PearlandMackenzie, | | | | | man, | 2020; | Chen et | al., 2020; | Feder | et | al., 2022) | | --- | --- | --- | --- | ---- | ----- | ------- | ---------- | ----- | --- | ---------- | 2018),underpinsthe“whatif”inquiriesessentialto haveenhancedthepredictiveaccuracy,robustness, humancognitionanddecision-makingacrosscrit- andexplainabilityofNLPmodels,recentprogress icalfieldssuchashealthcare,business,publicad- inLLMshasfurtherenrichedourabilitytocapture ministration,andscience(Gvozdenovic´etal.,2021; | | | | | nuanced | causal | dependencies | | (Liu | et | al., 2024a; | | --- | --- | --- | --- | ------- | ------ | ------------ | --- | ---- | --- | ----------- | Kyrimietal.,2025;KasirzadehandSmart,2021; Petronietal.,2019;Liangetal.,2024;Tarassow, *CorrespondingEmail:ruixiang.tang@rutgers.edu 2023;Ma,2024;Liuetal.,2024b). Theseadvance- mentsnotonlydemonstratesophisticatedreasoning cess,guidingLLMsthroughabduction,action,it- intasksrangingfromwritingtoprogrammingbut erativeprediction,andbacktrackingvalidationto also pave the way toward emulating human-like systematically formalize and explore counterfac- intelligenceandachievingartificialgeneralintelli- tual paths. This structured mechanism dynami- gence(LiandLi,2024;Alwin,2023;Sahota,2023; callyassessesthepromiseofeachinferencestep, Bubecketal.,2023). enabling reversion to more promising points and Despiterecentadvancements,progressincoun- ensuringlogicalconsistency,particularlyinlong- terfactual reasoning using LLMs remains con- chain causal dependencies. This systematic pro- strained by two primary challenges. First, there cesssubstantiallyimprovestheaccuracyofcoun- is currently no dedicated benchmark dataset for terfactualanalysis. ExperimentsonCounterBench rigorouslyevaluatingLLMs’performanceoncoun- demonstrate that CoIn achieves an accuracy of terfactualtasks,makingitdifficulttomeasurethe 89.9%,deliveringanearly20%improvementover models’ capacity to capture nuanced causal rela- Gemini-1.5-flashcomparedtoalternativestrategies tionships. Second,evenwithadvancedprompting (seeFigure1). Theframeworkofthisworkisillus- techniques,CausalCoT(Jinetal.,2023)andother tratedinFigure2. Thecontributionsofthiswork iterativemethods,LLMsoftenstruggletoproduce aresummarizedasfollows: logicallyconsistent,contextuallyappropriatecoun- terfactuals (Ma, 2024; Jin et al., 2023; Kıcıman • Webuildacomprehensivedataset,Counter- etal.,2023;Zecˇevic´ etal.,2023). Inresponse,this Bench. Thedatasetcontainsover1200long- paperfocusesontwokeyquestions: chaincomplexcounterfactualreasoningques- HowwelldoLLMshandlecounterfactualrea- tions. The dataset spans multiple difficulty soning? Theabsenceofastandardizedbenchmark levels,diversecausalgraphstructures,various datasethasimpededrigorousempiricalevaluation typesofcounterfactualquestions,andawide ofLLMs’capabilitiesincapturingintricatecausal range of nonsensical variant name combina- relationships within complex counterfactual rea- tions. soningtasks. Toaddressthis,wepresentCounter- Bench,acomprehensiveevaluationframeworkde- • WebenchmarkLLMswithvariousinference signedtoassesscounterfactualreasoningthrough strategies on CounterBench, and results re- 1.2Kquestionsencompassingvariousdomainsand vealthatmostexistingmodels(e.g.,GPT-4o reasoningtypes. Bysystematicallyevaluatingfive andDeepseek-V3)exhibitlimitedcapabilities keydimensions,itdemandsgenuinereasoningbe- inperformingcounterfactualinferencetasks. yondpatternrecognitionormemorizedresponses. Ourexperimentsexposenotableperformancelim- • WeproposeanovelreasoningparadigmCoIn itations in LLMs, even those equipped with ad- guides LLMs through abduction, action, it- vanced inference techniques. Most models like erative prediction, and backtracking valida- GPT-4oandDeepseek-V3achieveaccuracyofap- tiontosystematicallyformalizeandexplore proximately50%,equivalenttorandomguessing. counterfactual reasoning paths. It achieves Furthermore,ourevaluationofstate-of-the-artin- nearly90%accuracyonseveralstate-of-the- ference strategies shows only marginal improve- artLLMsevaluatedonCounterBench,repre- mentsoverbaselineperformanceformostmodels. sentinga20%improvementovertheprevious Themodelsconsistentlystrugglewithmaintaining bestbaseline. logicalcoherenceduringmulti-stepreasoningpro- cessesandaccuratelyhandlingcausalrelationships incomplexscenarios. 2 CounterBench How to improve LLMs’ counterfactual rea- soningabilities? Toadvancelargelanguagemod- Toevaluatethecounterfactualreasoningcapabili- els’counterfactualreasoningcapabilities,thispa- tiesofLLMs,weintroduceacomprehensivebench- per presents CoIn (Counterfactual Inference), a markingdatasetspecificallydesignedtomeasure novel approach that explicitly tackles the critical their ability to handle complex causal reasoning challengesofmulti-stepinference,whichremain tasks. This section details the structure of the unresolvedbypreviousmethods. CoInembedsa dataset,themethodologyforquerygeneration,and tailored search algorithm into the reasoning pro- thebenchmarkingresultsanalysis. CounterBench Dataset Benchmarking LLMs CoIn Reasoning Strategy | | Basic | | | Standard Strategy | | | | Extraction | | | --- | ----- | --- | --- | ----------------- | --- | --- | --- | ---------- | --- | without Instruction Joint Abduction and Intervention Action | | Nested | Human | | | Error | | | | | | --- | ------ | ------------ | --- | --------------- | -------- | --- | --- | --- | --- | | | | Verification | | Solver Strategy | Analysis | | | | | Forward Inference Conditonal | | Backdoor | | | CausalCoT Strategy | | | Back-tracking Validation | | | | --- | -------- | --- | --- | ------------------ | --- | --- | ------------------------ | --- | --- | Figure2: Illustrationoftheframework. WecreateCounterBench,adatasetfeaturingfivetypesofcounterfactual questions(basic,joint,conditional,nested,andbackdoor). Basedonthisdataset,webenchmarkstate-of-the-art LLMsusingvariousinferencestrategies,conductcomprehensiveerroranalysis,andproposeourCoInreasoning frameworkfeaturingsystematicinferencewithvalidationmechanisms. | Type | QueryTemplateExample | | | | | | Causalgraph | | | | ---- | -------------------- | --- | --- | --- | --- | --- | ----------- | --- | --- | WeknowthatXcausesV1,V1causesV2,V2causesV3,andV3causesV4, Basic V4causesV5,V5causesY.WouldYoccurifnotXinsteadofX? WeknowthatXcausesV1,V1causesV2,V2andV1togethercauseV3, | Joint | V3causesV4,V4andXtogethercauseV5,andV5causesY. | | | | | | | | | | ----- | ---------------------------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | WouldYoccurifnotXandnotV3? WeknowthatXcausesV1,V1causesV2,V2andV1togethercauseV3, V3causesV4,V4andV2togethercauseV5,andV5causesY. Nested AssumenotX,andbasedonthisassumption,furthersupposenotV4. WouldYoccur? WeknowthatXandV1togethercauseV2,V2causesV3,V3causesV4, | Conditional | V4causesV5,V5causesY.WeobservedV1.WouldYoccurifnotX | | | | | | | | | | ----------- | --------------------------------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | insteadofX? X WeknowthatV1causesX,XandV1togethercauseV2,V2causesV3, | Backdoor | | | | | | | | V2 V3 | V4 V5 Y | | -------- | --- | --- | --- | --- | --- | --- | --- | ----- | ------- | V3andXtogethercauseV4,V4causesV5,andV5causesY. V1 | | | | Table1: | IllustrativeCounterfactualQueryTypes | | | | | | | --- | --- | --- | ------- | ------------------------------------ | --- | --- | --- | --- | --- | 2.1 DatasetStructure BasicCounterfactual. Thebasiccounterfactual | | | | | | typeaddressessimple“what-if”scenarios. | | | | Inthis | | --- | --- | --- | --- | --- | -------------------------------------- | --- | --- | ----------------- | ------ | | | | | | | scenario,itisformalizedasY | | | (u),whichservesas | | x The dataset consists of two main components: a apotentialoutcomeexpression. Incausalreason- | | | | | | ing, potential | outcomes | refer | to the | hypothetical | | ------ | -------------- | ------- | --- | ------------- | -------------- | -------- | ----- | ------ | ------------ | | set of | counterfactual | queries | and | corresponding | | | | | | binary answers. Formally, the dataset is defined resultsobservedwhenavariableissettoapartic- as D := {(q ,a )|i = 1,2,...,N}, where each ular value (Holland, 1986). In Y (u) expression, | | i i | | | | | | | x | | | --- | --- | --- | --- | --- | --- | --- | ---------- | --- | --- | | | | | | | Y | | variable,x | | | q is a counterfactual query, and a ∈ {yes,no} is the outcome is the value consid- | i | | | | i | | | | | | | ---------- | ----------- | ------- | ---- | ------------ | ------- | ---------------- | --- | --------- | ------------- | | | | | | | ered in | the hypothetical | | scenario, | and u denotes | | represents | the correct | answer. | Each | query is de- | | | | | | rivedfromadeterministicStructuralCausalModel the context. To illustrate, consider a lawn irriga- (SCM)M = ⟨U,V,f⟩,whereU isthesetofexoge- tionsystemthatonlyactivateswhentheweatheris | | | | | | sunny and | the soil | is dry. | In this example, | Y (u) | | ------------------------------- | --- | --- | --- | ---------- | --------- | -------- | ------- | ---------------- | ----- | | nousvariableswithassignmentsu,V | | | | isthesetof | | | | | x | endogenousvariables,andf isthesetofstructural describesthesystem’spotentialbehaviorwhenthe equations(Pearl,2009). ForeachV ∈ V,wehave weatherconditionxisimposedwhilethesoilcondi- i | | | | | | tion(contextu)remainsunchanged. | | | | Consequently, | | --- | ----------- | ----------- | --- | ------------- | ------------------------------- | --- | --- | --- | ------------- | | V = | f (Pa(V ),U | ),wherePa(V | | ) ⊆ V denotes | | | | | | | i | i i | i | i | | | | | | | whenaskingwhetherthesystemwouldactivateif | the parents | of V i , | and U i refers | to | the subset of | | | | | | | ----------- | -------- | -------------- | --- | ------------- | --- | --- | --- | --- | --- | exogenousvariablesfromU thatdirectlyinfluence theweatherchangedtocloudy,therelevantcoun- the value of V . Intervening on a set of variables terfactualoutcomeisY cloudy (u). i X ⊆ V andsettingthemtoxmodifiesthemodelto M ,whichdeterministicallydefinesthevaluesof JointCounterfactual. Thistypeinvolvesacoun- x intervenedvariablesgivenu. Thedatasetincludes terfactualscenarioinwhichmultiplevariablesare fivetypesofcounterfactualqueries: set simultaneously. Formally, it is expressed as | | | | Standard | | | | CausalCoT | | | | --- | --- | --- | -------- | --- | --- | --- | --------- | --- | --- | Models Basic Cond. Joint NestedBack. Avg. Basic Cond. Joint NestedBack. Avg. GPT-3(Davinci-002) 56.8 50.2 48.8 51.6 52.5 51.9 51.2 41.9 51.2 51.6 50.5 49.3 GPT-3(Babbage-002) 50.0† 50.0† 50.0† 50.0† 47.5 49.6 3.6* 7.6* 1.2* 19.6* 18.5* 9.8* GPT-3.5 49.6 51.2 50.4 50.0 52.0 50.6 43.6 50.4 53.6 50.0 47.5 49.1 GPT-4omini 50.0† 50.0† 50.0† 50.0† 52.5 50.4 57.2 66.4 60.0 63.2 50.0 59.8 GPT-4o 50.4 54.4 50.4 54.8 54.0 52.8 80.4 72.4 80.8 81.6 60.5 75.8 Claude-3(Sonnet) 50.4 48.8 50.0 50.8 59.5 51.6 59.2 52.0 64.4 60.0 65.5 59.0 Claude-3.5(Haiku) 28.4 24.0 43.6 54.0 51.0 39.8 60.4 65.6 67.2 66.0 61.0 64.2 Gemini-1.5-flash 75.2 65.6 67.2 76.0 53.0 68.0 72.4 70.8 72.4 78.4 58.5 71.0 Gemini-1.5-flash-8b 50.0† 50.0† 50.0† 50.0† 52.5 50.4 66.8 67.2 65.2 65.2 58.5 64.8 Deepseek-V3 50.4 50.4 50.0 50.0 60.5 51.9 80.8 70.4 76.4 77.6 63.5 74.2 Table2: ModelaccuracyofstandardmethodandCausalCoTacrossdifferentreasoningcategories. Note: *The averageaccuracyisonly9.8%becausemostofresponsesarenot“Yes”or“No”but“incomprehensible”,which meansLLMcannotfollowinstructionofCausalCoTinstructionwelltoinfer. Moredetailswillbeexplainedinthe AppendixH.†indicatesthattheLLMpredictsallquestionsaseither“Yes”or“No”,leadingtoa50%accuracy. Y x,z (u),representingtheoutcomeY aftersetting observedassunny,thequeryevaluateswhetherthe X = xandZ = z. Forinstance,alawnirrigation lawn irrigation system still activate or not if the systemwillactivateiftheweatherissunny,butit sensordetectsmoistsoilinsteadofdryness. Here, alsorequiresdrysoilasatrigger. SupposeZ repre- Z = z representstheweatherisobservedassunny, sentstheweatherconditionandX representsthe whichisagivenconditionforreasoning. | soilmoisturecondition. | | Thequeryasksiftheirriga- | | | | | | | | | ---------------------- | --- | ------------------------ | --- | --- | --- | --- | --- | --- | --- | tionsystemwillactivatewhentheweatherchanges | to cloudy and | the sensor | detects | that the soil is | | | | | | | | --------------- | ----------------------------- | ------- | ---------------- | --- | --- | --- | --- | --- | --- | | moistmeanwhile. | Therelevantcounterfactualout- | | | | | | | | | comeisY (u). Thisscenarioexamines Backdoor Counterfactual. This type involves cloudy,moist thecombinedeffectofbothactionshappeningsi- counterfactual reasoning in the presence of back- | | | | | door paths | that create | confounding | | between | the | | --- | --- | --- | --- | ---------- | ----------- | ----------- | --- | ------- | --- | multaneously. treatmentvariableandtheoutcome(Pearl,2009). NestedCounterfactual. Nestedcounterfactual Formally, it addresses queries of the form Y x (u) | | | | | when there | exist | backdoor | paths | from X | to Y | | ------------------- | ------------ | --- | ------------- | ---------- | ----- | -------- | ----- | ------ | ---- | | involves sequential | dependencies | | between vari- | | | | | | | ables. This is represented as Y (u), where an throughconfounders. Insuchscenarios,thecausal Zx interventiononX affectsZ,whichinturnimpacts effect cannot be directly identified without con- | | | | | trolling | for the confounding | | variables | along | the | | --------------- | ------ | ------- | ---------------- | -------- | ------------------- | --- | --------- | ----- | --- | | Y. For example, | if the | weather | had been cloudy, | | | | | | | whichisacounterfactualweatherstate,andunder backdoor paths. For example, a manager consid- thisscenario,thesensordetectedmoistsoilinstead ersevaluatingwhetheranewmarketingcampaign of dry soil, would the irrigation system activate? would increase sales if it is implemented. How- Inthisframework,Z representstheweathercondi- ever,boththedecisiontolaunchthecampaignand tion,X isthesoilmoisturereading. Z dependson thesalesoutcomemightbeinfluencedbyseasonal thecounterfactualvalueofX throughthesystem’s demand patterns. Here, the backdoor path runs | | | | | from the | marketing | campaign | through | seasonal | | | --- | --- | --- | --- | -------- | --------- | -------- | ------- | -------- | --- | structuralcausalrelationships. | | | | | demand | to sales, | creating | a spurious | association. | | | --- | --- | --- | --- | ------ | --------- | -------- | ---------- | ------------ | --- | Conditional Counterfactual. This type intro- The counterfactual query “Would sales increase ducesobservedconditionsintothecounterfactual if we launched the campaign?” requires account- world (Pearl, 2009). Formally, it is written as ingforthisconfoundingbyeithercontrollingfor Y (u) | Z (u) = z, askinghowY wouldchange seasonaleffectsorusingotheridentificationstrate- | x x | | | | | | | | | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | ifX wassettoxwhileZ = z beingobservedas gies. Inourdataset,backdoorcounterfactualstest a condition. For example, a lawn irrigation sys- whether LLMs can distinguish between genuine temwillactivateiftheweatherissunny,butitalso causaleffectsandspuriouscorrelationswhenrea- requiresdrysoilasatrigger. Nowtheweatheris soningaboutalternativescenarios. 2.2 QueryGenerationandQuality 2022). Byintegratingasystematicderivationpro- | Assessment | | | | | | cess,includingcausalgraphextraction,querytype | | | | | | | | ---------- | -------- | --- | ---------- | ----------- | --- | --------------------------------------------- | --- | ---------------- | --- | --- | -------------- | --- | | | | | | | | classification, | | data collection, | | and | formalization, | | | Each query | consists | of | background | information | | | | | | | | | CausalCoTensuresrobustlogicalconsistencyand | and a specific | question. | | Table | 1 illustrates | how | | | | | | | | | -------------- | --------- | --- | ----- | ------------- | --- | --- | --- | --- | --- | --- | --- | --- | highreasoningaccuracy. samplesaregeneratedusingvariousdeterministic | | | | | | | EvaluationSettings. | | | Withinourevaluationframe- | | | | | ------------------------- | ----------- | ------ | ------------------- | ---------------- | --- | ------------------- | --- | -------------- | ------------------------- | ---- | ----- | -------- | | counterfactualquerytypes. | | | Thebackgroundiscon- | | | | | | | | | | | | | | | | | work, responses | | are classified | | into | three | distinct | | structed | with causal | graphs | and | story templates, | | | | | | | | | categories,“Yes”,“No”,and“Incomprehensible”. andvariablenamesarereplacedbynonsensical,ar- Thelatterencompassingresponsesthatareeither tificiallygeneratedwords(e.g.,“Kelp,”“Ziklo”)to ambiguousorlackclearmeaning,typicallymani- preventmodelsfromrelyingonmemorizedknowl- festingwhennoanswerisdetected,suchasNULL | edge. In | this way, | we | force LLMs | to engage | in | | | | | | | | | -------- | --------- | --- | ---------- | --------- | --- | ------- | ------- | ----- | ------- | ------ | ---------- | --- | | | | | | | | returns | or mere | query | echoes. | During | inference, | | causalreasoningratherthanusingpriorknowledge | | | | | | | wesetthetemperatureat0. | | | | Weemployinference | | | | -------------- | ----- | --- | ------- | ------------- | ---- | ----------------------- | --- | --- | --- | ----------------- | --- | --- | | in pretraining | data. | The | dataset | also features | bal- | | | | | | | | accuracyasourprimaryperformancemetric. | anced distributions | | in multiple | | dimensions, | with | | | | | | | | | ------------------- | --- | ----------- | --- | ----------- | ---- | --- | --- | --- | --- | --- | --- | --- | binaryresponsesevenlysplitbetween50%“Yes” 3.1 ExperimentalResults and50%“No.”Thisbalanceextendsacrossdiffer- AsshowninTable2,theresultsindicatethatwith- entquestiontypesanddifficultylevels,ensuringa outspecificinstructions,mostLLMsstrugglewith uniformresponsedistributionwithineachcategory. counterfactualreasoning,performingnobetterthan Thedatasetconsistsof1,200questions,categorized | | | | | | | randomguessingintermsofaccuracy. | | | | | Specifically, | | | --- | --- | --- | --- | --- | --- | -------------------------------- | --- | --- | --- | --- | ------------- | --- | intofivedistincttypes,witheachtypecontaining | | | | | | | for model | GPT-4o | mini, | we | observed | consistent | | | ------------------ | --- | ------------------------ | --- | --- | --- | ----------- | ------ | ------ | ----- | -------- | ---------- | --- | | 200or250questions. | | Withineachtype,thereisan | | | | | | | | | | | | | | | | | | predictions | of | either | “Yes” | or “No,” | resulting | in | equaldistributionofanswers,comprising100“Yes” | | | | | | | a 50.0% | accuracy | in | the first | four | kinds | of ques- | | --------- | ------- | ------------- | --------- | ------- | ---------- | -------------------------------------------- | -------- | --- | --------- | ---- | ----- | -------- | | responses | and 100 | “No” | responses | or 125 | “Yes” | | | | | | | | | | | | | | | tions. Amongalltestedmodels,Gemini-1.5-flash | | | | | | | | and 125 | “No”. | Additionally, | the | dataset | is strati- | | | | | | | | achievedthehighestbaselineperformancewithan | fied based | on five | levels | of difficulty, | determined | | | | | | | | | | ---------- | ------- | ------ | -------------- | ---------- | --- | -------- | --------- | -------- | --- | --- | --------- | --- | | | | | | | | accuracy | of 68.0%. | Although | | the | CausalCoT | ap- | bythenumberofeventspresentineachquestion, proachisdesignedtoenhancethecausalreasoning | rangingfrom5to9. | | Eachdifficultylevelincludes | | | | | | | | | | | | ---------------- | --- | --------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | capabilitiesofLLMs,ourempiricalfindingssug- 240questions,maintainingabalanceddistribution gestthatitdoesnotsignificantlyimprovetheirper- | ofanswerswith120“Yes”and120“No”. | | | | | Wealso | | | | | | | | | -------------------------------- | --- | --- | --- | --- | ------ | --------------------------------------- | --- | --- | --- | --- | --- | ---- | | | | | | | | formanceincounterfactualreasoningtasks. | | | | | | Most | conductahumanevaluationonthesequeries,with modelsshowedminimalornoimprovement,asex- furtherdetailsprovidedinAppendixB. | | | | | | | emplifiedbyGPT-3.5Turbo. | | | | Thebestperformance | | | | --- | --- | --- | --- | --- | --- | ------------------------ | ------------- | --- | --- | ------------------ | -------- | --- | | | | | | | | model in | the CausalCoT | | is | GPT-4o, | achieved | an | 3 BenchmarkingLLMsonCounterBench accuracyofonly75.8%. Weconductedcomprehensiveexperimentstosys- ErrorAnalysis. Tosystematicallyanalyzethelim- tematically evaluate the performance of current itations of existing approaches, we conducted an LLMs on counterfactual reasoning tasks, demon- error analysis on responses generated by Causal- stratingtheircapabilitiesusingstate-of-the-artrea- CoT. Our analysis focuses on three key compo- soningtechniques. | | | | | | | nents: causal | | data collection, | | inference | | process, | | --- | --- | --- | --- | --- | --- | ------------- | --- | ---------------- | --- | --------- | --- | -------- | Models. The tested LLM models include GPT- and conclusion derivation. Through careful ex- 3.5 turbo, GPT-4o, GPT-4o mini, Davinci-002, aminationofeachcomponent,weidentifiedthree Babbage-002(OpenAI,2024),Claude3.5Haiku, primarycategoriesoferrors: Wrongcausalrela- Claude 3 Sonnet (Anthropic, 2024), Deepseek- tionships: This error occurs when LLMs cannot V3 (DeepSeek, 2024) and Gemini-1.5-Flash and constructaccuratecausalgraphsorextractknown Gemini-1.5-Flash-8B(Google,2024). valuesfrombackgroundinformation. Wronginfer- ReasoningStrategies. Inourbaselineevaluations, enceprocess: ThishappenswhenLLMs,despite we employed two distinct reasoning strategies to correctlyidentifyingcausalrelationships,makein- assess these models. The first relied on standard correctpredictionsofthetargeteventY. Wrong prompting methods without specialized instruc- conclusion: ThistypeoferrorariseswhenLLMs tions. The second used the advanced CausalCoT reach contradictory final answers, even after cor- approach (Jin et al., 2023), an extension of the rectlycomputingthevalueofY. Thedistribution Chain-of-Thoughtpromptingparadigm(Weietal., of these error categories is illustrated in Figure 3. Notably, 86% of errors occur in the inference Each phase serves a specific purpose in ensuring process,revealingthatevenwithwell-constructed accurate counterfactual reasoning, and they pro- causal graphs, LLMs struggle significantly with vide a robust methodology for handling complex derivingaccuratepredictionsthroughreasoning. causaldependenciestogether. Below,wedescribe eachphaseindetail,explainingitsroleandhowit contributestotheoverallprocess. 4.1 Extraction Inthefirstphase,wefocusonsystematicallygath- 12% | | 2% | | ering all | relevant information | | explicitly | stated | in | | --- | --- | --- | -------------- | -------------------- | -------------- | ---------- | ---------------- | ----- | | | | | the scenario. | The | process | begins | with construct- | | | | | | ing the causal | graph | by identifying | | relationships | | | | | | between | events and | representing | | them in a | clear | | | | | “event 1 | → event | 2” format, | | which eliminates | | | | | 86% | potential | ambiguities. | Next, | we | collect the | given | valuesforeachvariablefrombothbackgroundin- Wrong Inference | | Wrong Relations | | formationandquestions,wherethesevaluesindi- | | | | | | | --- | --------------- | --- | ------------------------------------------- | --- | --- | --- | --- | --- | Wrong Conclusion | | | | catewhetherspecificeventsoccurornot. | | | | Crucially, | | | --- | --- | --- | ------------------------------------ | --- | --- | --- | ---------- | --- | thisphasemaintainsstrictadherencetoexplicitly Figure3: ErrorAnalysisofCausalCoT. stated information, avoiding any unsupported in- ferencesorassumptionsinfavorofarigorousand unbiaseddatacollectionprocess. 4 ProposedReasoningStrategy 4.2 Abduction Asdiscussedinprevioussection,theprimarychal- lenge for large language models is to minimize Thisphasefocusesoninferringtheposteriorcon- incorrectinferences,whichareamajorsourceof straintsovertheexogenousnoisevariables,equiv- errors. Toaddressthischallenge,weproposeCoIn alently, constraints over parent assignments that (Counterfactual Inference), a systematic reason- make the observed factual world consistent with ingframeworkthatguideslargelanguagemodels thestructuralequations. Foreachobservedvariable throughstructuredproblem-solvinginsteadofre- V with value v , we invert its structural equa- obs lyingonintuitiveshortcutsormemorizedpatterns. tionV := f (Parents(V),U )toobtaineithera | | | | | V | | V | | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | Our approach transforms counterfactual queries unique solution for U V or a feasible set over U V into a five-phase algorithmic process, mirroring giventheparents. Indeterministiclogicalmodels, how humans naturally approach “what if” ques- thisisoftenconvenientlycarriedoutbydeducing tions (Sel et al., 2023): first understanding what parent assignments that must hold for v to be obs actuallyhappened,thenimaginingthealternative true. Theresultingvaluesarestoredasthefactual scenario,systematicallyworkingthroughthecon- world knowledge base and will be held fixed in sequences, andfinallydouble-checkingthelogic. subsequentInterventionActionandForwardInfer- Thisstructuredapproachsignificantlyreducesrea- ence,ensuringcounterfactualsareevaluatedinthe | soningerrorsbybreakingdowncomplexproblems | | | sameworld. | | | | | | | ------------------------------------------ | --- | --- | ---------- | --- | --- | --- | --- | --- | | intomanageablestepswithbuilt-invalidation. | | An | | | | | | | 4.3 InterventionAction exampleofproposedparadigmisprovidedinAp- | pendixA. | | | In this phase, | the framework | | applies | the counter- | | | -------- | --- | --- | -------------- | ------------- | --- | ------- | ------------ | --- | TheCoInframeworkconsistsoffivekeyphases: factualinterventionsdescribedinthequery. This Extraction: ExtractCounterfactualInformation involvesmodifyingtheoriginalsetofcausalrules from the given natural language facts. Abduc- byreplacingtheequationsfortheintervenedvari- tion: Infer the underlying conditions from ob- ableswithconstantvalues,resultinginanupdated servedfacts;InterventionAction: Applythehy- setofrules. Theinterventionsareincorporatedinto potheticalchangesspecifiedinthequery;Forward theknowledgebase,effectivelymakingaprecise Inference: Systematicallytracethroughthecausal alterationtothecausalgraph. Thisphasecaptures consequences;Back-trackingValidation: Verify thecore“whatif”elementofthequery,allowing thelogicalconsistencyoftheentirereasoningchain. theframeworktosimulatehypotheticalworldsin a controlled way. It focuses on specific changes, Model Standard CausalCoTSolver Ours whichstreamlinestheexplorationbylimitingthe | | | | | | | | GPT-3(Davinci-002) | | | 51.9 | 49.3 | 50.1 | 49.6 | | --- | --- | --- | --- | --- | --- | --- | ------------------ | --- | --- | ---- | ---- | ---- | ---- | search to paths directly affected by the interven- GPT-3 (Babbage- 49.6 9.8 47.9 45.8 002) | tion, similar | | to how | efficient | | searches | eliminate | | | | | | | | | ------------- | --- | ------ | --------- | --- | -------- | --------- | ---------- | --- | --- | ---- | ---- | ---- | ---- | | | | | | | | | GPT-4omini | | | 50.4 | 59.8 | 47.2 | 79.9 | unnecessarybranches. | | | | | | | | GPT-4o | | | 52.8 | 75.8 | 51.4 | 89.4 | | -------------------- | --- | --- | --- | --- | --- | --- | ----------------- | --- | --- | ---- | ---- | ---- | ---- | | | | | | | | | GPT-3.5turbo | | | 50.6 | 49.1 | 49.6 | 58.9 | | 4.4 ForwardInference | | | | | | | Claude-3(Sonnet) | | | 51.6 | 59.0 | 51.8 | 89.8 | | | | | | | | | Claude-3.5(Haiku) | | | 39.8 | 64.2 | 48.3 | 79.1 | Duringthisiterativephase,theframeworkpredicts Gemini-1.5-flash 68.0 71.0 52.8 89.9 | | | | | | | | Gemini-1.5-flash-8b | | | 50.4 | 64.8 | 50.3 | 83.9 | | --- | --- | --- | --- | --- | --- | --- | ------------------- | --- | --- | ---- | ---- | ---- | ---- | valuesforunobservedvariablesbyselectingnodes | | | | | | | | Deepseek-V3 | | | 51.9 | 74.2 | 49.3 | 91.8 | | --- | --- | --- | --- | --- | --- | --- | ----------- | --- | --- | ---- | ---- | ---- | ---- | inthecausalgraphwhoseparentvariablesareal- ready known in the knowledge base and evaluat- Table3: ModelaccuracyonCounterBench. Wereport ing their updated equations. Beginning with the theaverageaccuracyforfourinferencestrategies: Stan- intervenedvariablesandtheinferrednoiseterms, dard,CausalCoT,Solver,andCoIn. | it gradually | computes | | the | effects | on | downstream | | | | | | | | | ------------------------------------------ | -------- | --- | --- | ------- | --- | ---------- | --- | --- | --- | --- | --- | --- | --- | | variablesuntilitreachesthetargetvariableY. | | | | | | Ifa | | | | | | | | approachcombinesLLMswithStructuralCausal node’svaluecannotbecalculatedduetomissingin- Model(SCM)tools(Pearl,2009)forcausalinfer- formationaboutitsparents,theframeworkchooses ence. ThestudyintroducesCausalTool,asuiteof | another | suitable | node | and | continues | | the process | | | | | | | | | ------- | -------- | ---- | --- | --------- | --- | ----------- | --- | --- | --- | --- | --- | --- | --- | 10inferencetoolsdesignedforvariouscausaltasks. | until Y | is determined. | | This | forward | | progression | | | | | | | | | ------- | -------------- | --- | ---- | ------- | --- | ----------- | --- | --- | --- | --- | --- | --- | --- | ItleveragesLLMstoclassifycausalquestions,ex- mimicsadepth-firstexplorationofcausalchains, tractcausalgraphsandformalizeddata,androute enablingtheframeworktodynamicallyconstruct | | | | | | | | them to | the appropriate | | tools | for | inference, | with | | ---------- | --------- | ----------- | --------- | ------------ | --- | ----------- | --------------------------------- | --------------- | --- | ----- | --- | ---------- | ------ | | and assess | potential | | outcomes. | | By | focusing on | | | | | | | | | | | | | | | | thefinalanswergeneratedbytheLLM.1 | | | | | | During | | nodes that | can | be computed | | immediately, | | it nav- | | | | | | | | inference,thetemperatureissettozero. | igates the | dependency | | graph | efficiently, | | steering | | | | | | | | | ---------- | ------------ | --- | ----- | ------------ | --------- | -------- | --- | --- | --- | --- | --- | --- | --- | | clear of | unproductive | | paths | and | promoting | a me- | | | | | | | | 5.2 MainResult | thodical | advancement | | toward | | the solution. | The | | | | | | | | | -------- | ----------- | --- | ------ | --- | ------------- | --- | --- | --- | --- | --- | --- | --- | --- | detailsofthisphaseisintheAppendixG. The comprehensive performance comparison | | | | | | | | across all | models | is | presented | in | Table | 3. Our | | --- | --- | --- | --- | --- | --- | --- | ---------- | ------ | --- | --------- | --- | ----- | ------ | 4.5 Back-trackingValidation approachdemonstratesnotableimprovementsover | | | | | | | | existing | methods | across | | the model | spectrum, | | | ---------- | --- | -------- | --------- | --- | ------ | --------- | -------- | ------- | ------ | --- | --------- | --------- | --- | | To confirm | | that the | predicted | | values | are logi- | | | | | | | | callyconsistent,thisfinalphaseretracesthesteps with particularly noteworthy performance gains | | | | | | | | achieved | by smaller | | language | models, | including | | | ------- | ------------- | --- | ---- | --- | ------------ | --- | -------- | ---------- | --- | -------- | ------- | --------- | --- | | through | the knowledge | | base | and | re-evaluates | the | | | | | | | | GPT-4omini,Claude-3.5Haiku,andGemini-1.5- equationforeachnon-noisevariableusingthepre- | | | | | | | | flash-8b. | Forinstance,ourmethodenablesGPT-4o | | | | | | | ------------- | --- | -------------------------------- | --- | --- | --- | --- | --------- | ---------------------------------- | --- | --- | --- | --- | --- | | dictedvalues. | | ForeverysuchvariableV,itrecalcu- | | | | | | | | | | | | minitoachieveanaccuracyof79.9%,surpassing latestheexpectedvaluebasedontheupdatedequa- theperformanceofseverallargermodelswithout tionandchecksifitmatchesthepreviouslystored | | | | | | | | CoInenhancement. | | AsdetailedinTable4,taking | | | | | | --- | --- | --- | --- | --- | --- | --- | ---------------- | --- | ------------------------- | --- | --- | --- | --- | value. Ifanymismatchoccurs,theframeworksig- nalsanerror,highlightingapotentialissueinthe GPT-4ominiasanexample,CoInachievessuperior performanceacrossallfivetypesofcounterfactual | earlierreasoningthatmayneedreevaluation. | | | | | | This | | | | | | | | | ---------------------------------------- | ------ | --- | ------------ | --- | ------- | ------- | ---------- | ---- | ------------ | --- | ------ | ------- | ------ | | | | | | | | | questions, | with | particularly | | better | results | on ba- | | validation | serves | as | a protective | | measure | against | | | | | | | | sicquestionscomparedtomorecomplexvariants. | errors that | could | accumulate | | during | | the process, | | | | | | | | | ----------- | ----- | ---------- | --- | ------ | --- | ------------ | ------------- | ------------------------------ | --- | --- | --- | --- | --- | | | | | | | | | Additionally, | state-of-the-artLLMssuchasGPT- | | | | | | akintoretracingapathtoconfirmitsvalidity. 4o,Gemini-1.5-flash,andDeepseek-V3achievere- | 5 Experiments | | | | | | | markableaccuracyapproachingorexceeding90% | | | | | | | | ------------- | --- | --- | --- | --- | --- | --- | ----------------------------------------- | --- | --- | --- | --- | ----------- | --- | | | | | | | | | whenaugmentedwithourmethod. | | | | | TakenGPT-4o | | 5.1 ExperimentSetup as example, Our strategy improves the accuracy WeadoptedthesameLLMsasmentionedinSec- ofthemodelfrom75.8%to89.4%,demonstrating tion3forourexperiments. Toestablishbaselines, CoIn’seffectivenessinguidingLLMsthroughal- | | | | | | | | gorithmtoexplorepathsstep-by-step. | | | | | Theresults | | | --- | --- | --- | --- | --- | --- | --- | ---------------------------------- | --- | --- | --- | --- | ---------- | --- | weimplementedbothCausalCoT(Jinetal.,2023) | andstandardsolverstrategies. | | | | Thelatterintegrates | | | | | | | | | | | ---------------------------- | --- | --- | --- | ------------------- | --- | --- | --- | --- | --- | --- | --- | --- | --- | 1SincethesourcecodeforCausalToolisnotpubliclyavail- externaltoolsintothechain-of-thoughtprocess,as able,were-implementeditscounterfactualinferenceproce- described in (Hua et al., 2024). Specifically, this durebasedondescriptionsintheoriginalpaper. Methods Basic Cond. Joint Nested Back. Avg. terfactualreasoningtasks. Standard 50.0 50.0 50.0 50.0 52.5 50.4 90 CausalCoT 57.2 66.4 60.0 63.2 50.0 59.8 Solver 35.2 54.4 50.4 50.0 45.5 47.2 80 Ours 82.8 79.2 80.0 80.4 76.5 79.9 70 60 Table 4: Accuracy of GPT-4o mini across five query 50 typesintheCounterBench. 40 30 indicatethatcontemporaryLLMs,whenequipped 20 withourstrategy,caneffectivelyresolvemostfor- 10 malcomplexcounterfactualproblems. Thedetails 0 gemini-1.5 gemini-1.5 gemini-1.5-8bgemini-1.5-8b ofallperformanceofourresultsarepresentedin anti- common- anti- common- commonsense sense commonsense sense Appendix C. Moreover, in Appendix D, we con- ductederroranalysis. Theanalysisrevealsasub- stantial reduction in errors of inference process. Moreover,wealsoexaminetheimpactofcomplex causal relationships on outcomes. We found that accuracy decreases as the number of variants in- creasesforCausalCoTandCoIn. Thedetailsare showninthenextsection. 5.3 ValidatingGeneralizationAbility Inthissection, weevaluatethegeneralizationca- pabilityofCoInusingtheCLADDERdataset(Jin etal.,2023). CLADDERisadatasetfocusonthe causalreasoningquestions. WeutilizetheCLAD- DERdatasettodetermineiftheproposedmethod canbeextendedbeyondtheCounterBenchdataset. Unlikeourdataset,whichfocusesonformalrules, CLADDER includes examples that utilize com- monsenseknowledgeratherthancausalinference abilitiestoanswerqueries. Specifically,itencom- passesbothcommonsenseandanti-commonsense scenarios, allowing us to explore whether CoIn remains effective under the influence of pretrain- ingknowledgeinLLMs. Weconductexperiments on the counterfactual subset of CLADDER. We applied the Gemini-1.5 and Gemini-1.5-8b mod- elstobothcommonsenseandanti-commonsense queries,withtheresultsdisplayedinFigure4and detailed examples in Appendix E. These results demonstratethatperformanceisconsistentlystable across different reasoning paradigms, suggesting that pretraining knowledge has a limited impact ontheCoIncounterfactualreasoningcapabilities. Furthermore,ourmethodachievesanaccuracyof 78.98%,outperformingbothCausalCoTat64.77% andtheStandardmethodat64.20%. Thisperfor- manceunderscoresCoIn’sgeneralizabilityandits potential for broader application in various coun- )%( ycaruccA Standard CausalCoT Ours Figure 4: Accuracy comparison between Standard, CoIn, and CausalCoT method in Anti-commonsense andCommonsenseDataset. 6 RelatedWork CounterfactualReasoning. Counterfactualrea- soning explores how outcomes change when cer- tainvariablesarealteredfromtheirhistoricalstates. InStructuralCausalModels(SCMs),Pearl’s(Pearl, 2009) “surgery” and do-calculus provide system- aticwaystoinferinterventionoutcomes,highlight- ing deep causal knowledge required for accurate inference. Counterfactuals can be deterministic or probabilistic: deterministic settings yield pre- dictable outcomes from given conditions, while probabilistic models incorporate inherent uncer- tainties. These methods have gained traction in domains like social sciences, where they assess alternativepolicyoutcomesandstudycausalmech- anismsinobservationaldata(Morgan,2015),and inmedicine,wheretheyenablepersonalizedtreat- mentanddecisionsupport(Johanssonetal.,2016; Shalitetal.,2017;Louizosetal.,2017;Yoonetal., 2018). Inartificialintelligence,counterfactualrea- soning is crucial for interpretability and fairness, enablingmodelstogeneratealternativescenarios andassessdecision-makingrobustness. Although recent efforts extend counterfactual reasoning to LLMs(Jinetal.,2023),significantchallengesper- sist,particularlyregardingcomplexvariablerela- tionships in high-dimensional text data. Conse- quently,bridgingthegapbetweentextualcomplex- ityandrobustcausalinferenceremainsafocalpoint forfutureresearch. LLMs in Counterfactual Learning. With the rapidevolutionofLLMs,theresearchcommunity hasincreasinglyfocusedontheirabilitytoperform Sébastien Bubeck, Varun Chandrasekaran, Ronen El- causalinference(Zhangetal.,2023;Ashwanietal., dan, Johannes Gehrke, Eric Horvitz, Ece Kamar, | | | | | | | | Peter Lee, | Yin Tat | Lee, | Yuanzhi | Li, Scott | Lund- | | --- | --- | --- | --- | --- | --- | --- | ---------- | ------- | ---- | ------- | --------- | ----- | 2024). AprominentexampleisCausalAgent,an | | | | | | | | berg,etal.2023. | | Sparksofartificialgeneralintelli- | | | | | --- | --- | --- | --- | --- | --- | --- | --------------- | --- | --------------------------------- | --- | --- | --- | agent-basedLLMframeworkthatmergesanLLM | | | | | | | | gence: Earlyexperimentswithgpt-4. | | | | arXivpreprint | | | ----------- | ----- | --- | ------- | ----- | ---- | ------- | --------------------------------- | --- | --- | --- | ------------- | --- | | with causal | tools | for | complex | tasks | (Han | et al., | | | | | | | arXiv:2303.12712. | 2024). | While it | excels | at identifying | | causal | asso- | | | | | | | | ------ | -------- | ------ | -------------- | --- | ------ | ----- | --- | --- | --- | --- | --- | --- | IviChatzi,NinaCorveloBenz,EleniStraitouri,Stratis | ciations | and conducting | | interventions, | | | it largely | | | | | | | | -------- | -------------- | --- | -------------- | --- | --- | ---------- | --------------------------------------- | --- | --- | --- | --- | ----- | | | | | | | | | Tsirtsis,andManuelGomez-Rodriguez.2024. | | | | | Coun- | omitscounterfactualreasoning,limitingitsapplica- terfactualtokengenerationinlargelanguagemodels. bilitytomoreadvancedscenarios. Currentefforts arXivpreprintarXiv:2409.17027. | to integrate | counterfactual | | | reasoning | into | LLMs | | | | | | | | ------------ | -------------- | --- | --- | --------- | ---- | ---- | --- | --- | --- | --- | --- | --- | HuigangChen,TotteHarinen,Jeong-YoonLee,Mike | typically | follow | two | paths. | First, | commonsense- | | | | | | | | | --------- | ------ | --- | ------ | ------ | ------------ | --- | --------- | ------ | ----- | --------------- | --- | ------ | | | | | | | | | Yung, and | Zhenyu | Zhao. | 2020. Causalml: | | Python | basedapproachesleveragebackgroundknowledge | | | | | | | | packageforcausalmachinelearning. | | | | arXivpreprint | | | ----------- | ---------- | ------- | --------- | ----------- | ------- | -------- | -------------------------------- | ----------- | --- | ------------------- | ------------- | --- | | to imagine | scenarios | | that defy | established | | facts | arXiv:2002.11631. | | | | | | | (Ning et | al., 2024; | Chatzi | et | al., | 2024; | Musi and | | | | | | | | | | | | | | | DeepSeek.2024. | DeepSeek: | | AI-PoweredSearchEn- | | | | Palmieri, | 2024; | Vicuna, | 2023), | such | as | positing | | | | | | | | | | | | | | | gine. Accessed: | 2025-02-15. | | | | | | alternative | historical | | outcomes. | | Second, | graph- | | | | | | | based methods employ formal causal graphs and AmirFeder,KatherineAKeith,EmaadManzoor,Reid Pryzant,DhanyaSridhar,ZachWood-Doughty,Jacob externalPythonpackagesforcomputations,asseen Eisenstein,JustinGrimmer,RoiReichart,MargaretE | inCausalTool(Huaetal.,2024). | | | | | Althoughthese | | | | | | | | | ---------------------------- | --- | --- | --- | --- | ------------- | --- | ------------------ | --- | ---------------------------- | --- | --- | --- | | | | | | | | | Roberts,etal.2022. | | Causalinferenceinnaturallan- | | | | methodseffectivelyincorporatestructuredcausal | | | | | | | | guageprocessing: | | Estimation,prediction,interpreta- | | | | | --- | --- | --- | --- | --- | --- | --- | ---------------- | --- | --------------------------------- | --- | --- | --- | information, they often offload key calculations tionandbeyond. TransactionsoftheAssociationfor ComputationalLinguistics,10:1138–1158. outsidetheLLM. | | | | | | | | Google. 2024. | Gemini. | | https://gemini.google. | | | | ------------ | --- | --- | --- | --- | --- | --- | --------------- | ----------- | --- | ---------------------- | --- | --- | | 7 Conclusion | | | | | | | com/. Accessed: | 2025-01-06. | | | | | Inthiswork,wedevelopandextendCounterBench, IanDGow,DavidFLarcker,andPeterCReiss.2016. acounterfactualreasoningdatasetwithfiveprob- Causalinferenceinaccountingresearch. Journalof AccountingResearch,54(2):477–523. | lem types | for LLM | evaluation. | | | Our findings | re- | | | | | | | | --------- | ------- | ----------- | --- | --- | ------------ | --- | --- | --- | --- | --- | --- | --- | vealthatmostLLMsperformnear-randomly,with Emilia Gvozdenovic´, Lucio Malvisi, Elisa Cinconze, StijnVansteelandt,PhoebeNakanwagi,Emmanuel state-of-the-artmethodsshowingminimalimprove- ment. To address these challenges, we propose Aris,andDominiqueRosillon.2021. Causalinfer- enceconceptsappliedtothreeobservationalstudies | CoIn, a | reasoning | paradigm | | inspired | by | formal | | | | | | | | ------- | --------- | -------- | --- | -------- | --- | ------ | --------------------------------- | --- | --- | --- | ---------- | --- | | | | | | | | | inthecontextofvaccinedevelopment: | | | | fromtheory | | causalinferenceprinciplesandplanningstrategies. to practice. BMC Medical Research Methodology, | CoInguidesLLMsthroughiterativethinkingand | | | | | | | 21:1–10. | | | | | | | ----------------------------------------- | --- | --- | --- | --- | --- | --- | -------- | --- | --- | --- | --- | --- | backtrackingtoexplorereasoningpathsmoreeffec- KairongHan,KunKuang,ZiyuZhao,JunjianYe,and tively. Ourapproachsignificantlyenhancescoun- | | | | | | | | FeiWu.2024. | Causalagentbasedonlargelanguage | | | | | | --- | --- | --- | --- | --- | --- | --- | ----------- | ------------------------------- | --- | --- | --- | --- | terfactualreasoningcapabilitiesofLLMs. model. arXivpreprintarXiv:2408.06849. | | | | | | | | Paul W Holland. | 1986. | Statistics | | and causal | infer- | | --- | --- | --- | --- | --- | --- | --- | --------------- | ----- | ---------- | --- | ---------- | ------ | ence. JournaloftheAmericanstatisticalAssociation, References 81(396):945–960. | Alwin. | 2023. Understanding | | | causal | ai: Bridging | the | | | | | | | | ----------------------- | ------------------- | --- | --- | ---------- | ------------ | -------- | ------------- | ------- | ----- | ------- | --------- | ------ | | | | | | | | | Zhenyang Hua, | Shuyue | Xing, | Huixing | Jiang, | Chen | | gap between | correlation | | and | causation. | | https:// | | | | | | | | | | | | | | | Wei, and | Xiaojie | Wang. | 2024. | Improving | causal | | www.alwin.io/causal-ai. | | | | Accessed: | 2025-01-06. | | | | | | | | inferenceoflargelanguagemodelswithscmtools. | | | | | | | | In CCF International | | Conference | | on Natural | Lan- | | --- | --- | --- | --- | --- | --- | --- | -------------------- | --- | ---------- | --- | ---------- | ---- | Anthropic. 2024. Claude. https://www.anthropic. guage Processing and Chinese Computing, pages | com/api. | Accessed: | | 2025-01-06. | | | | | | | | | | | -------- | --------- | --- | ----------- | --- | --- | --- | --- | --- | --- | --- | --- | --- | 3–14.Springer. Swagata Ashwani, Kshiteesh Hegde, Nishith Reddy Zhijing Jin, Yuen Chen, Felix Leeb, Luigi Gresele, Mannuru, Dushyant Singh Sengar, Mayank Jindal, Ojasv Kamal, LYU Zhiheng, Kevin Blin, Fer- Krishna Chaitanya Rao Kathala, Dishant Banga, nando Gonzalez Adauto, Max Kleiman-Weiner, Vinija Jain, and Aman Chadha. 2024. Cause and Mrinmaya Sachan, et al. 2023. Cladder: Assess- effect: Canlargelanguagemodelstrulyunderstand ingcausalreasoninginlanguagemodels. InThirty- causality? InProceedingsoftheAAAISymposium seventhconferenceonneuralinformationprocessing | Series,volume4,pages2–9. | | | | | | | systems. | | | | | | | ------------------------ | --- | --- | --- | --- | --- | --- | -------- | --- | --- | --- | --- | --- | FredrikJohansson,UriShalit,andDavidSontag.2016. JingMa.2024. Causalinferencewithlargelanguage Learningrepresentationsforcounterfactualinference. model: Asurvey. arXivpreprintarXiv:2409.09822. | In International | conference | | on machine | learning, | | | | | | | | ---------------- | ---------- | --- | ---------- | --------- | --- | --- | --- | --- | --- | --- | pages3020–3029.PMLR. SL Morgan. 2015. Counterfactuals and causal infer- ence. CambridgeUniversityPress. | AtoosaKasirzadehandAndrewSmart.2021. | | | | Theuse | | | | | | | | ------------------------------------ | --- | --- | --- | ------ | --- | --- | --- | --- | --- | --- | and misuse of counterfactuals in ethical machine Elena Musi and Rudi Palmieri. 2024. The fallacy of learning. InProceedingsofthe2021ACMConfer- explainablegenerativeai: evidencefromargumenta- enceonFairness,Accountability,andTransparency, tivepromptingintwodomains. InCEURWorkshop | pages228–236. | | | | | Proceedings,volume3769,pages59–69. | | | | | | | ------------- | --- | --- | --- | --- | ---------------------------------- | ----- | ---- | ------ | ----- | ---------- | | | | | | | Xuefei Ning, | Zinan | Lin, | Zixuan | Zhou, | Zifu Wang, | EmreKıcıman,RobertNess,AmitSharma,andChen- haoTan.2023. Causalreasoningandlargelanguage HuazhongYang,andYuWang.2024. Skeleton-of- models: Openinganewfrontierforcausality. arXiv thought: Promptingllmsforefficientparallelgener- preprintarXiv:2305.00050. ation. In The Twelfth International Conference on LearningRepresentations. LisaKoonce,KarenKNelson,andCatherineMShake- speare. 2011. Judging the relevance of fair value OpenAI.2024. Models. https://platform.openai. for financial instruments. The Accounting Review, com/docs/models. Accessed: 2025-01-06. 86(6):2075–2098. | | | | | | Judea Pearl. | 2009. | Causality. | | Cambridge | university | | ----------------------------------------------- | --- | -------- | --------------- | --- | ------------ | ----- | ---------- | --- | --------- | ---------- | | ParthasarathyKrishnamurthyandAnuradhaSivaraman. | | | | | press. | | | | | | | 2002. Counterfactual | | thinking | and advertising | re- | | | | | | | sponses. JournalofConsumerResearch,28(4):650– JudeaPearlandDanaMackenzie.2018. Thebookof | | | | | | why: | the new | science | of cause | and | effect. | | ---- | --- | --- | --- | --- | ---- | ------- | ------- | -------- | --- | ------- | | 658. | | | | | | | | | | Basic | books. | Evangelia Kyrimi, | Somayyeh | | Mossadegh, | Jared M | | | | | | | | ----------------- | -------- | --- | ---------- | ------- | --- | --- | --- | --- | --- | --- | Wohlgemut, RebeccaSStoner, NigelRMTai, and Fabio Petroni, Tim Rocktäschel, Patrick Lewis, An- WilliamMarsh.2025. Counterfactualreasoningus- tonBakhtin,YuxiangWu,AlexanderHMiller,and ingcausalbayesiannetworksasahealthcaregover- SebastianRiedel.2019. Languagemodelsasknowl- arXivpreprintarXiv:1909.01066. | nancetool. InternationalJournalofMedicalInfor- | | | | | edgebases? | | | | | | | ---------------------------------------------- | --- | --- | --- | --- | ---------- | --- | --- | --- | --- | --- | matics,193:105681. | | | | | | Neil Sahota. | | 2023. | Causal | | ai: Bridging | | --- | --- | --- | --- | --- | ------------ | --- | ----- | ------ | --- | ------------ | Jia Li and Xiang Li. 2024. Relation-first modeling the gap between correlation and causation. paradigmforcausalrepresentationlearningtoward https://www.neilsahota.com. Accessed: 2025-01- | thedevelopmentofagi. | | Preprint,arXiv:2307.16387. | | | 06. | | | | | | | -------------------- | --- | -------------------------- | --- | --- | --- | --- | --- | --- | --- | --- | Weixin Liang, Yaohui Zhang, Zhengxuan Wu, Haley Bilgehan Sel, Ahmad Al-Tawaha, Vanshaj Khattar, Lepp,WenlongJi,XuandongZhao,HanchengCao, Ruoxi Jia, and Ming Jin. 2023. Algorithm of ShengLiu,SiyuHe,ZhiHuang,etal.2024. Mapping thoughts: Enhancing exploration of ideas in large theincreasinguseofllmsinscientificpapers. arXiv languagemodels. arXivpreprintarXiv:2308.10379. preprintarXiv:2404.01268. | | | | | | Uri Shalit, | Fredrik | D Johansson, | | and | David Sontag. | | --- | --- | --- | --- | --- | ----------- | ------- | ------------ | --- | --- | ------------- | JinxinLiu,ShulinCao,JiaxinShi,TingjianZhang,Lun- 2017. Estimating individual treatment effect: gen- yiuNie,LinmeiHu,LeiHou,andJuanziLi.2024a. eralizationboundsandalgorithms. InInternational Howproficientarelargelanguagemodelsinformal conferenceonmachinelearning,pages3076–3085. | languages? anin-depthinsightforknowledgebase | | | | | PMLR. | | | | | | | -------------------------------------------- | --- | -------------------------- | --- | --- | ----- | --- | --- | --- | --- | --- | | questionanswering. | | InFindingsoftheAssociation | | | | | | | | | forComputationalLinguisticsACL2024,pages792– Amit Sharma and Emre Kiciman. 2020. Dowhy: | 815. | | | | | An end-to-end | | library | for causal | inference. | arXiv | | ---- | --- | --- | --- | --- | ------------- | --- | ------- | ---------- | ---------- | ----- | preprintarXiv:2011.04216. XiaoyuLiu,PaihengXu,JundaWu,JiaxinYuan,Yifan Yang,YuhangZhou,FuxiaoLiu,TianruiGuan,Hao- ArturTarassow.2023. Thepotentialofllmsforcoding liangWang,TongYu,etal.2024b. Largelanguage withlow-resourceanddomain-specificprogramming modelsandcausalinferenceincollaboration: Acom- languages. arXivpreprintarXiv:2307.13018. | prehensivesurvey. | arXivpreprintarXiv:2403.09606. | | | | | | | | | | | ----------------- | ------------------------------ | --- | --- | --- | ------- | ----- | ------- | -------------- | --- | ----------- | | | | | | | Vicuna. | 2023. | Vicuna: | An open-source | | chatbot im- | MassimoLoiandMargaridaRodrigues.2012. Anote pressingGPT-4with90%*ChatGPTquality. https: ontheimpactevaluationofpublicpolicies: thecoun- //vicuna.lmsys.org/. Accessed: 2023. terfactualanalysis. JasonWei,XuezhiWang,DaleSchuurmans,Maarten ChristosLouizos,UriShalit,JorisMMooij,DavidSon- Bosma,FeiXia,EdChi,QuocVLe,DennyZhou, tag,RichardZemel,andMaxWelling.2017. Causal etal.2022. Chain-of-thoughtpromptingelicitsrea- effectinferencewithdeeplatent-variablemodels. Ad- soninginlargelanguagemodels. Advancesinneural vancesinneuralinformationprocessingsystems,30. informationprocessingsystems,35:24824–24837. Jinsung Yoon, James Jordon, and Mihaela Van lemsofsufficientcomplexitytoeffectivelyevaluate Der Schaar. 2018. Ganite: Estimation of individ- counterfactualreasoningabilitiesofmodels. The ualizedtreatmenteffectsusinggenerativeadversarial | | | | | | inter-annotator | | agreement | | between | the | two PhD | | --- | --- | --- | --- | --- | --------------- | --- | --------- | --- | ------- | --- | ------- | nets. InInternationalconferenceonlearningrepre- annotatorswasapproximately95%,demonstrating sentations. | | | | | | strongconsistencyintheirjudgments. | | | | | | Beforethe | | --- | --- | --- | --- | --- | ---------------------------------- | --- | --- | --- | --- | --- | --------- | MatejZecˇevic´,MoritzWillig,DevendraSinghDhami, | | | | | | full annotation | | began, | both | annotators | | were pro- | | ------------------------- | --- | -------------- | --- | ----- | --------------- | --- | ------ | ---- | ---------- | --- | --------- | | andKristianKersting.2023. | | Causalparrots: | | Large | | | | | | | | videdwithasetofpracticeexamplesanddetailed languagemodelsmaytalkcausalitybutarenotcausal. arXivpreprintarXiv:2308.13067. reasoning guidelines. They participated in a cali- brationphaseinvolvingdiscussionandalignment | Cheng Zhang, | Stefan Bauer, | Paul | Bennett, | Jiangfeng | | | | | | | | | ------------ | ------------- | ---- | -------- | --------- | --- | --- | --- | --- | --- | --- | --- | onannotationstandards,ensuringasharedunder- | Gao, Wenbo | Gong, Agrin | Hilmkil, | Joel | Jennings, | | | | | | | | | ---------- | ----------- | -------- | ---- | --------- | --- | --- | --- | --- | --- | --- | --- | ChaoMa,TomMinka,NickPawlowski,etal.2023. standingofthetask. Thispreparatorystephelped establishconsistencyandreliabilityacrossthefull | Understanding | causality | with large | language | mod- | | | | | | | | | ---------------- | ------------------ | ---------- | -------- | -------- | -------- | --- | --- | --- | --- | --- | --- | | els: Feasibility | and opportunities. | | arXiv | preprint | dataset. | | | | | | | arXiv:2304.05524. C.MoreExperimentDetails Appendix | | | | | | Table 5 | presents | a | comprehensive | | performance | | | --- | --- | --- | --- | --- | ------- | -------- | --- | ------------- | --- | ----------- | --- | A.MethodPromptDesign | | | | | | comparison | between | | the Solver | | method | and our | | --- | --- | --- | --- | --- | ---------- | ------- | --- | ---------- | --- | ------ | ------- | Ourpromptdesign,asillustratedinFigure5,has CoInapproachacrossvariousmodelsontheCoun- | | | | | | terBenchdataset. | | Ouranalysisrevealsseveralsig- | | | | | | --- | --- | --- | --- | --- | ---------------- | --- | ----------------------------- | --- | --- | --- | --- | beencarefullystructuredtooptimizetheinteraction betweenouralgorithmandLargeLanguageMod- nificantpatternsinmodelperformanceacrossdif- els (LLMs). At its core, the design incorporates ferent counterfactual reasoning tasks. First and multiplefew-shotexamplesthatserveascompre- foremost,CoIndemonstratesconsistentsuperiority hensive demonstrations of our algorithm’s opera- over the Solver method across all task categories tionalframework. Theseexamplesarestrategically andmodelarchitectures. Thisimprovementispar- selected to showcase various reasoning patterns ticularlypronouncedinnewermodelarchitectures, withstate-of-the-artmodelslikeGPT-4o,Claude-3 andproblem-solvingapproaches,providingLLMs witharobustfoundationforunderstandingthealgo- (Sonnet), and Gemini-1.5-flash showing remark- rithm’smethodology. Withintheprompt,wehave ableperformancegains. Notably,GPT-4oachieves perfectaccuracy(100.0%)onbasicreasoningtasks meticulouslydetailedstep-by-stepinstructions,cre- atingaclearandstructuredinformationflowthat underourmethod. Deepseek-V3demonstratesex- guides the model through the reasoning process. ceptionalconsistencywithhighperformanceacross ThisstructuredapproachleveragesLLMs’inherent all task types. Its accuracy comes to 99.6% for | | | | | | basic tasks | and | maintaining | | above | 90% | average | | --- | --- | --- | --- | --- | ----------- | --- | ----------- | --- | ----- | --- | ------- | in-contextlearningcapabilities,enablingthemto systematically explore solution paths rather than accuracy. Theperformancedistributionacrossdif- relyingonintuitiveresponses. Thisapproachnot ferenttasktypesrevealsinterestingpatterns. Basic | | | | | | reasoning | tasks | consistently | | yield | the | highest ac- | | ------------- | ----------- | ------- | -------- | ---- | --------- | ----- | ------------ | --- | ----- | --- | ----------- | | only enhances | the model’s | ability | to break | down | | | | | | | | problems and evaluate paths efficiently but also curacy under our method, particularly evident in reducingthelikelihoodofgeneratingerroneousin- largermodels. However,thiscategoryalsoexhibits termediatestepsandimprovingreasoningstability. themostsignificantperformancevariationacross | | | | | | different | model | architectures, | | suggesting | | that ba- | | --- | --- | --- | --- | --- | --------- | ----- | -------------- | --- | ---------- | --- | -------- | siccounterfactualreasoningcapabilitiesarehighly B.HumanEvaluation | | | | | | sensitive | to model | scale | and | architecture. | | In con- | | --- | --- | --- | --- | --- | --------- | -------- | ----- | --- | ------------- | --- | ------- | Toassessthequalityofourdataset,weaskedtwo trast, joint reasoning tasks show relatively stable Ph.D. students with expertise in causal inference performance across different models, indicating toanswer200randomlyselectedquestions. They thatthesecapabilitiesmaybemorefundamentally achieved an average accuracy of 97.75% and re- tiedtothereasoningframeworkratherthanmodel quiredfourhourstocompletethem,reflectingthe size. Moreover,weobserveaclearcorrelationbe- substantial cognitive effort involved. According tweenmodelrecencyandperformance,withnewer toparticipantfeedback,thesequestionsdemanded modelslikeGPT-4oandClaude-3(Sonnet)achiev- carefulconsiderationofmultiplecausalfactorsand ingsignificantlyhigheraverageaccuracies(89.8%, explicitcounterfactualreasoningsteps. Thesefind- 89.4%)comparedtotheirpredecessors. Thistrend ings demonstrate that our dataset presents prob- holdstrueacrossalltasktypes,thoughthemagni- tudeofimprovementvariesbycategory. Thecon- in Figure 12, the approach breaks down the rea- sistent performance improvements across newer soningprocessintodistinctphases: counterfactual modelarchitecturessuggestthatrecentadvancesin informationcollection,andsystematicexploration languagemodeldevelopmenthaveenhancedtheir ofinferencepaths. Thisstructureddecomposition capacity for structured counterfactual reasoning contrastssharplywiththestandardapproachshown whencombinedwithourmethodology. inFigure9,whichprovidesminimalguidancefor | | | | | | | navigatingthereasoningprocess. | Throughexplicit | | | --- | --- | --- | --- | --- | --- | ------------------------------ | --------------- | --- | D.ErrorAnalysiscomparisonbetweenOur variablemappingandsystematicpathexploration, MethodandCausalCoT | | | | | | | CoIn enables | models to systematically | evaluate | | --- | --- | --- | --- | --- | --- | ------------ | ------------------------ | -------- | By randomly sampling 50 instances and catego- possible causal chains, leading to more reliable rizingerrorsintothreedistincttypes, wrongrela- andtraceableinferenceoutcomes. Thekeyreason tions, wrong inferences, and wrong conclusions, why only CoIn arrived at the correct answer lies we systematically evaluated the model’s perfor- initsexplicitstructuredreasoningprocess,which mance. Theanalysisdemonstratednotablediffer- ensuresasystematicandrobustapproachtocoun- ences in the relative distribution of errors, with terfactual inference. Unlike CausalCoT and the inference-related errors decreasing from 86% to standardmethod,CoInemploysastep-by-stepal- 46%. This significant reduction combining with gorithmicframeworkthatsystematicallyprocesses thediminishederrorquantitysubstantiatesanen- causaldependencies,preventingshortcutreasoning | hanced | counterfactual | inference | capability. | | Con- | andheuristicerrors. | | | | -------------------- | -------------- | --------- | ----------- | -------- | ------ | ------------------- | --- | --- | | currently, | we also | observed | a relative | increase | in | | | | | relationship-related | | errors, | 12% to | 50%. | Conse- | | | | quently,theoveralleffectofourstrategyispositive asthenumberoferrorscomedownnotably. E.CLADDERDatasetExample TwoexamplesaregeneratedfromtheCLADDER | dataset. | ItisshowninFigure | | 7andFigure8,are | | | | | | | -------- | ----------------- | --- | --------------- | --- | --- | --- | --- | --- | designedtoevaluateamodel’sabilitytodistinguish betweencommonsenseandanti-commonsenserea- | soningincounterfactualscenarios. | | | | Commonsense | | | | | | -------------------------------- | ------- | ------ | ------------- | ----------- | ----- | --- | --- | --- | | reasoning | follows | causal | relationships | that | align | | | | G.ForwardInferenceAlgorithm | with human | intuition | and | everyday | knowledge, | | | | | | ---------- | --------- | --- | -------- | ---------- | --- | --- | --- | --- | makingiteasierformodelstoinferoutcomesbased | onfamiliarpatterns. | | Incontrast,anti-commonsense | | | | | | | | ------------------- | -------- | --------------------------- | ---------- | ---- | ------- | --- | --- | --- | | reasoning | presents | causal | structures | that | contra- | | | | dictintuitiveexpectations,requiringmodelstorely solelyonexplicitlyprovidedcausalrelationships | rather than | prior | knowledge. | By testing | | both rea- | | | | | ----------- | ----- | ---------- | ---------- | --- | --------- | --- | --- | --- | soningparadigms,theseexamplesassesswhether ThepartisforwardinferenceAlgorithm,thecore amodelcanaccuratelydifferentiatebetweenintu- methodologicalcomponentthatsystematicallyap- itiveandcounterintuitivecausalstructures,ensur- plies gathered information from previous Extrac- ingthatreal-worldbiasesdonotinterferewithits tion, Abduction, andActionstepstoevaluatethe counterfactualreasoningabilities. target event. The framework employs iterative counterfactualreasoningtoprogressivelyexplore F.Answerofourmethods andinfereventvalues,ultimatelydeterminingthe Here,weprovideacomparativeanalysisbetween targetoutcome. Moreover,throughintegratedeval- our CoIn method, CausalCoT and the standard uation and backtracking mechanisms, the frame- approach. The results are shown in Figure 9, work enables models to systematically optimize Figure 11 and Figure 12. In results, our CoIn theirreasoningpathsandimprovereasoningaccu- method introduces a structured, step-by-step rea- racy. Thisprocessfollowsanalgorithmicstrategy soning framework that systematically addresses designedtosystematicallydeterminethecounter- | complex | counterfactual | | scenarios. | As illustrated | | factualoutcome. | | | | ------- | -------------- | --- | ---------- | -------------- | --- | --------------- | --- | --- | | | | | | | Solver | | | | | | Ours | | | --- | --- | --- | --- | --- | ------ | --- | --- | --- | --- | --- | ---- | --- | Models Basic Cond. Joint Nested Back. Avg. Basic Cond. Joint Nested Back. Avg. GPT-3(Davinci-002) 50.4 50.0 50.4 50.0 49.5 50.1 51.2 48.4 48 51.6 48.5 49.6 GPT-3 (Babbage- 44.8 48.8 48.4 48.0 50.0 47.9 45.2 46.4 46 45.6 46.0 45.8 002) GPT-3.5turbo 50.0 50.0 50.0 50.0 47.5 49.6 65.2 54.0 61.2 63.2 49.0 58.9 GPT-4omini 35.2 54.4 50.4 50.0 45.5 47.2 82.8 79.2 80.0 80.4 76.5 79.9 GPT-4o 45.6 60.8 51.6 50.4 48.0 51.4 100.0 89.6 90.0 88.4 76.5 89.4 Claude-3(Sonnet) 55.6 58.0 43.6 50.4 51.5 51.8 97.6 84.8 90.0 94.0 80.5 89.8 Claude-3.5(Haiku) 49.2 49.2 48.8 49.6 43.5 48.3 88.0 76.8 81.2 83.2 63.0 79.1 Gemini-1.5-flash 54.8 58.8 50.8 50.8 48.0 52.8 98.4 89.6 90.0 94.0 74.5 89.9 Gemini-1.5-flash-8b 52.4 50.8 49.2 51.2 47.0 50.3 94.8 80.0 88.0 88.4 64.5 83.9 Deepseek-V3 51.2 48.0 48.4 50.4 48.0 49.3 99.6 90.0 90.4 94.0 83.0 91.8 Table5: PerformanceofSolversandOurCoInAcrossDifferentReasoningCategories Algorithm1ForwardInferenceAlgorithm anerroranalysis,categorizingtheerrorsintothree | | | | | | | | distincttypes: | BlankResponse: | | | Thisoccurswhen | | | --- | --- | --- | --- | --- | --- | --- | -------------- | -------------- | --- | --- | -------------- | --- | Input: | | R: | | | | | | the model | produces | an | empty | response. | Repeat- | | --- | --- | --- | --- | --- | --- | --- | --------- | -------- | --- | ----- | --------- | ------- | 1: Asetofcausalgraphrelations E: Asetofevents ing Error: This occurs when the model’s infer- 2: | | N: | | | | | | enceisonlyrepeatingthequestionsandgivenin- | | | | | | | --- | --- | --- | --- | --- | --- | --- | ------------------------------------------ | --- | --- | --- | --- | --- | 3: Asetof{event,value}pairs | | | | | | | | formations. | The | example | is | shown | in Figure 13. | | --- | --- | --- | --- | --- | --- | --- | ----------- | --- | ------- | --- | ----- | ------------- | Output: Y: targeteventY Type Mismatch: It represents that the answer is 4: 5: procedure REASONING(R,E,N) expectedtobeabinaryvalue“yes”or“no”. The | | | | | E\N.events | | | responsereturnsalistwithmultiple“yes”or“no”. | | | | | | | --- | ------ | ------------- | ---------- | ---------- | --- | --- | -------------------------------------------- | --- | --- | --- | --- | --- | | 6: | K | ← RANDOM(E),E | | ∈ | | | | | | | | | | | whileK | ∈/ | N.eventsdo | | | | TheexampleisshowninFigure14. | | | | | | 7: | | | forallrelationr | | ∈ Rdo | | | | | | | | | | --- | --- | --------------- | --- | ----- | --- | --- | --- | --- | --- | --- | --- | --- | 8: INFER(r,N) | 9: | | if | ∃ | | → value(K) | | | | | | | | | --- | --- | --- | --- | --- | ---------- | --- | --- | --- | --- | --- | --- | --- | 1.20 then | | | | N ← N | ∪{K,value(K)} | | | | | | | | | | --- | --- | --- | ----- | ------------- | --- | --- | ---- | --- | --- | --- | --- | --- | | 10: | | | | | | | 1.00 | | | | | | 0.18 | 11: | | | ifK = Y | then | | | | | | 0.19 | | | | --- | ---------- | --- | -------------- | ----------- | --- | --- | ---- | ---- | ---- | ---- | ---- | ---- | | | | | | | | | 0.80 | | | | | 0.32 | | | | | | | | | | 0.10 | 0.46 | 0.12 | 0.43 | | | 12: | | | returnvalue(K) | | | | | | | | | | | | | | else | | | | 0.60 | | | | | | | 13: | | | | | | | | | | | | 0.24 | | 14: | | | K ← | RANDOM(E),E | | ∈ | | | | | | | | | | | | | | | 0.40 | | 0.30 | | | | | | E\N.events | | | | | | | 0.72 | | 0.69 | 0.42 | | | | | | | | | | 0.20 | | | | | 0.45 | | 15: | | | endif | | | | | | 0.24 | | | | 0.14 | | | endif | | | | | 0.00 | | | | | | | --- | --- | ------ | --- | --- | --- | --- | ---- | -------------- | ----- | --------------- | ------------- | ---- | | 16: | | | | | | | | Basic | Cond. | Joint | Nested | Avg. | | 17: | | endfor | | | | | | Blank Response | | Repeating Error | Type Mismatch | | 18: endwhile endprocedure | 19: | | | | | | | Figure10: | ErrorAnalysisforBabbage-002inCausal- | | | | | | --- | --- | --- | --- | --- | --- | --- | --------- | ------------------------------------ | --- | --- | --- | --- | CoT | H.BenchmarkResultAnalysis: | | | | CausalCoT | | | | | | | | | | -------------------------- | --- | --- | --- | --------- | --- | --- | --- | --- | --- | --- | --- | --- | WecomparedtheresponsesgeneratedbyBabbage- | 002 | under | the CausalCoT | | framework | with | those | | | | | | | | -------------- | ----- | ------------- | ----------------------------- | --------- | ---- | ----- | --- | --- | --- | --- | --- | --- | | fromotherLLMs. | | | Theresultsrevealthatthemajor- | | | | | | | | | | ityofBabbage-002’sresponseswerenonsensical whenfollowingtheCausalCoTinstructions,result- | ing | in a remarkably | | low accuracy | | of just | 8.0%. | | | | | | | | -------- | --------------- | -------- | ------------ | ---------- | ------- | ------- | --- | --- | --- | --- | --- | --- | | A deeper | | analysis | showed | that 82.4% | of | the re- | | | | | | | sponseswereclassifiedasincomprehensible,which explainswhytheaccuracywasnotablylowerthan | random | guessing. | | The following | | chart provides | | | | | | | | | ------ | --------- | --- | ------------- | --- | -------------- | --- | --- | --- | --- | --- | --- | --- | Question: —————- User: Imagineaself-contained,hypotheticalworldwithonlythefollowingconditions,andwithout any unmentioned factors or causal relationships: Pexu has a direct effect on Kraz. Not Kraz and Pexuhasadirecteffectonhwax. Hwaxhasadirecteffectonrukz. Krazisunobserved. Weknow thatpexuandnotkrazcauseshwax. pexucauseskraz. hwaxcausesrukz. Weobservedanindividual ishwax. Wouldanindividualisrukzifnotkrazinsteadofkraz? Pleasefollowtheinstructionsstep bystep. Step1. Extractthecausalgraph: Identifythecausalgraphthatdepictstherelationshipsin thescenario. LetX=kraz;V2=pexu;V3=hwax;Y=rukz. Thediagramshouldsimplyconsistof edgesdenotedin"var1→var2"format,separatedbycommas. IfyougetV1→NotV2forexample, youneedtoreplaceintoV1→V2. Step2. AbductionandActionforinformationcollection: Collect all the directly given information into given values set. 1 means given observed in question or observed. 0meansgivennotinquestion. Donotassumeorinferothervariablesvaluesbyrelations. Then, describe relations about how multiple variables influence another variable; it can result in AND,OR,orNOT.Step3. InferandPredicttheYbyinformationstepbystep. Step4. Basedonthe resultfromtheStep3,derivethefinalanswer. Thereisanidentifiableanswer. Assistant: Step 1) Extract the causal graph: Identify the causal graph that depicts the relationships in the scenario. LetV2=kraz; X=pexu; V3=hwax; Y=rukz. ThecausalgraphisV2→V3, V2→X, X→V3,V3→Y. Step2)AbductionandActionforinformationcollection: Allgivenvalues: V2=1,andX=0(by abduction,V2occurs),X=0(byaction,Xnotoccur);Relations: Y:(NOTV2)ANDV3,V3: X ANDNOTV2. Step3)Tryingapromisingfirstoperation: 1. V3=notXandV2=1 -Y=V3=1→foundit! Backtrackingthesolution: Step3a: V3=notXandV2=1 Step3b: Y=V3=1 Step4)SincetheresultfortheYis1,theoverallanswertothequestionisyes. —————- User: Imagineaself-contained,hypotheticalworldwithonlythefollowingconditions,andwithout anyunmentionedfactorsorcausalrelationships: VankhasadirecteffectonPraf,Prafhasadirect effect on Scud, Scud has a direct effect on Wrenk, Wrenk and Scud have direct effects on Yobb, YobbhasadirecteffectonGlim,GlimandScudhavedirecteffectsonSpog,andSpoghasadirect effectonKlep. VankBern(0.6). WeobservedWrenk. WeknowthatVankcausesPraf,Prafcauses Scud,ScudcausesWrenk,WrenkandScudtogethercauseYobb,YobbcausesGlim,GlimandScud together cause Spog, and Spog causes Klep. Vank Bern(0.6). We observed Wrenk. Would Klep occur ifnot Prafinstead of Praf? Step 1. Extract thecausal graph: Identifythe causalgraph that depictstherelationshipsinthescenario. LetX=Ziklo;V1=Blaf;V2=Trune;V3=Vork;V4= Sline;V5=Frim;V6=Qado;Y=Lumbo. Thediagramshouldsimplyconsistofedgesdenoted in"var1→var2"format, separatedbycommas. IfyougetV1→NotV2forexample, youneed toreplaceintoV1→V2. Step2. AbductionandActionforinformationcollection: Collectallthe directlygiveninformationintogivenvaluesset. 1meansgivenobservedinquestionorobserved. 0 means given not in question. Do not assume or infer other variables values by relations. Then, describerelationsabouthowmultiplevariablesinfluenceanothervariable;itcanresultinAND,OR, orNOT.Step3. InferandPredicttheYbyinformationstepbystep. Step4. Basedontheresultfrom theStep3,derivethefinalanswer. Thereisanidentifiableanswer. Assistant: Figure5: ThepromptdesignexampleofCoIn 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% | | Wrong Relations | Wrong Inference | | Wrong Conclusion | | | | | | | | --------- | --------------- | --------------- | ---- | ---------------- | --- | --- | --- | --- | --- | --- | | | | CausalCoT | Ours | | | | | | | | | Figure 6: | Error Analysis | comparison | | between Our | | | | | | | MethodandCausalCoT. | | | | | | Question: | Imagine | | a self-contained, | | hy- | | --- | --- | --- | --- | --- | --------- | ------- | --- | ----------------- | --- | --- | Question: Imagineaself-contained,hypothet- pothetical world with only the following icalworldwithonlythefollowingconditions, | | | | | | conditions, | and | without | any | unmentioned | | | --- | --- | --- | --- | --- | ----------- | --- | ------- | --- | ----------- | --- | andwithoutanyunmentionedfactorsorcausal | | | | | | factors | or causal | relationships: | | Nuv | has a | | --- | --- | --- | --- | --- | ------- | --------- | -------------- | --- | --- | ----- | relationships: Having a brother has a direct direct effect on Splee, Blen and Druk, not | effectonroom. | | Thecandlehasadirecteffect | | | | | | | | | | ------------- | --------------------------- | ------------------------- | --- | --- | ------------- | -------- | ------ | -------- | ------- | ----- | | | | | | | Druk has | a direct | effect | on Plog, | Plog | has a | | onroom. | Weknowthathavingabrotherand | | | | | | | | | | | | | | | | direct effect | on | Skrim, | Skrim | or Druk | has | candle with wax causes dark room. We ob- a direct effect on Zimb, Zimb has a direct servedthecandlehaswax. Wouldtheroomis effectonYurd,andYurdhasadirecteffecton darkifnothavingabrotherinsteadofhaving Wrox. WeknowthatNuvcausesSplee,Blen abrother? andDruk,notDrukcausesPlog,Plogcauses Start your answer with “Yes” or “No”, fol- Skrim, Skrim or Druk causes Zimb, Zimb lowedbyadditionalreasoningorevidenceto | | | | | | causesYurd,andYurdcausesWrox. | | | | | Would | | --- | --- | --- | --- | --- | ----------------------------- | --- | --- | --- | --- | ----- | supportyourexplanation. WroxoccurifnotNuvandnotSplee? | | | | | | Answer: | No. | | | | | | --------- | ----------------------- | --- | --------------- | ----- | -------- | -------------------------- | --- | --- | --- | --- | | Figure7: | Anti-commonsenseExample | | | | | | | | | | | | | | | | Figure9: | StandardInstructionExample | | | | | | Question: | Imagine | a | self-contained, | hypo- | | | | | | | theticalworldwithonlythefollowingcondi- | tions, and | without | any | unmentioned | factors | | | | | | | | ----------------------- | -------------------------- | ----------------- | --------------- | ---------- | --- | --- | --- | --- | --- | --- | | orcausalrelationships: | | | Themanintheroom | | | | | | | | | hasadirecteffectonroom. | | | Thecandlehasa | | | | | | | | | directeffectonroom. | | Weknowthatblowing | | | | | | | | | | out the | candle | and candle | with | wax causes | | | | | | | | darkroom. | Weobservedthecandlehaswax. | | | | | | | | | | Wouldtheroomisdarkifnotblowingoutthe candleinsteadofblowingoutthecandle? | Start your | answer | with | “Yes” | or “No”, fol- | | | | | | | | ---------- | ------ | ---- | ----- | ------------- | --- | --- | --- | --- | --- | --- | lowedbyadditionalreasoningorevidenceto supportyourexplanation. Figure8: CommonsenseExample Question: Imagine a self-contained, hypothetical world with only the following conditions, and withoutanyunmentionedfactorsorcausalrelationships: NuvhasadirecteffectonSplee,Blenand Druk,notDrukhasadirecteffectonPlog,PloghasadirecteffectonSkrim,SkrimorDrukhasa directeffectonZimb,ZimbhasadirecteffectonYurd,andYurdhasadirecteffectonWrox. We knowthatNuvcausesSplee,BlenandDruk,notDrukcausesPlog,PlogcausesSkrim,SkrimorDruk causesZimb,ZimbcausesYurd,andYurdcausesWrox. WouldWroxoccurifnotNuvandnotSplee? Answer: Toaddresstheproblem,wewillfollowthestepsoutlined: Step1: Extractthecausalgraph Basedonthegivenconditions,thecausalgraphcanberepresentedasfollows: -Nuv→Splee-Nuv→Blen-Nuv→Druk-Druk→Plog-Plog→Skrim-Skrim→Zimb-Druk →Zimb-Zimb→Yurd-Yurd→Wrox Step2: Determinethequerytype The question asks whether Wrox would occur if not Nuv and not Splee. This is a counterfactual query because it involves reasoning about what would happen under a hypothetical scenario that differsfromtheobservedworld. Thequerytypeis: “counterfactual” Step3: Formalizethequery Theformalexpressionforthecounterfactualqueryis: Wrox Nuv=0,Splee=0 ThisnotationrepresentsthevalueofWroxinthecounterfactualworldwhereNuvandSpleedonot occur. Step4: Gatherallrelevantdata Fromtheproblemstatement,wehavethefollowingdata: -Nuv=0(notNuv)-Splee=0(notSplee) Step5: Deducetheestimandusingcausalinference Todeducetheestimand,weneedtoconsiderthecausalpathwaysfromNuvandSpleetoWrox: 1. Nuv directly affects Splee, Blen, and Druk. 2. Druk affects Plog, which affects Skrim, which affectsZimb,whichaffectsYurd,whichaffectsWrox. 3. DrukalsodirectlyaffectsZimb. 4. Zimb affectsYurd,whichaffectsWrox. GiventhatNuv=0,Splee=0,weneedtodetermineifWroxcanstilloccur: -SinceNuv=0,Splee=0,Blen=0,andDruk=0. -Druk=0impliesPlog=0andZimb=0(since DrukaffectsbothPlogandZimb). -Plog=0impliesSkrim=0. -Skrim=0andDruk=0imply Zimb=0. -Zimb=0impliesYurd=0. -Yurd=0impliesWrox=0. Thus,thecounterfactualscenarioresultsinWrox=0. Step6: Calculatetheestimand Giventhedeductionsabove,thecounterfactualvalueofWroxwhenNuv=0andSplee=0is: Wrox = 0 Nuv=0,Splee=0 Therefore,WroxwouldnotoccurifnotNuvandnotSplee. Figure11: CausalCoTInstructionExample Question: Imagine a self-contained, hypothetical world with only the following conditions, and withoutanyunmentionedfactorsorcausalrelationships: NuvhasadirecteffectonSplee,Blenand Druk,notDrukhasadirecteffectonPlog,PloghasadirecteffectonSkrim,SkrimorDrukhasa directeffectonZimb,ZimbhasadirecteffectonYurd,andYurdhasadirecteffectonWrox. We knowthatNuvcausesSplee,BlenandDruk,notDrukcausesPlog,PlogcausesSkrim,SkrimorDruk causesZimb,ZimbcausesYurd,andYurdcausesWrox. WouldWroxoccurifnotNuvandnotSplee? Answer: Step1)Extractthecausalgraph: Identifythecausalgraphthatdepictstherelationshipsin thescenario. LetX=Nuv;V1=Splee;V2=Blen;V3=Druk;V4=Plog;V5=Skrim;V6=Zimb; V7=Yurd;Y=Wrox. Thecausalgraphis: X→V1,X→V2,X→V3,V3→V4,V4→V5,V5→V6,V3→V6,V6→V7,V7→Y. Step2)AbductionandActionforinformationcollection: Allgivenvalues: X=0(notNuv),V1=0 (notSplee). Relations: -V1: X -V2: X -V3: X -V4: NOTV3 -V5: V4 -V6: V5ORV3 -V7: V6 -Y:V7 Step3)Adoptthefollowingalgorithmtogettheresult: InfertheYbyinformationstepbystep. 1. V1=X=0 2. V2=X=0 3. V3=X=0 4. V4=NOTV3=1 5. V5=V4=1 6. V6=V5ORV3=1OR0=1 7. V7=V6=1 8. Y=V7=1 Step 4) Conclude the final answer: Since the result for Y (Wrox) is 1, the overall answer to the questionisyes. WroxwouldoccurifnotNuvandnotSplee. Figure12: CoInInstructionExample Response GlentcausesRazz,RazzandGlenttogethercausePex,PexcausesZurn,ZurncausesMelf,andMelf andRazztogethercauseZlim. WouldZlimoccurifnotGlentinsteadofGlent? B:GlentcausesRazz,RazzandGlenttogethercausePex,PexcausesZurn,ZurncausesMelf,and MelfandRazztogethercauseZlim. WouldZlimoccurifnotGlentinsteadofGlent? C:GlentcausesRazz,RazzandGlenttogethercausePex,PexcausesZurn,ZurncausesMelf,and MelfandRazztogethercauseZlim. WouldZlimoccurifnotGlentinsteadofGlent? D:GlentcausesRazz,RazzandGlenttogethercausePex,PexcausesZurn,ZurncausesMelf,and MelfandRazztogethercauseZlim. WouldZlimoccurifnotGlentinsteadofGlent? E:GlentcausesRazz,RazzandGlenttogethercausePex,PexcausesZurn,ZurncausesMelf,and MelfandRazztogethercauseZlim. WouldZlimoccurifnotGlentinsteadofGlent? F:GlentcausesRazz,RazzandGlenttogethercausePex,PexcausesZurn,ZurncausesMelf,and MelfandRazztogethercauseZlim. WouldZlimoccurifnotGlentinsteadofGlent? G:GlentcausesRazz,RazzandGlenttogethercausePex,PexcausesZurn,ZurncausesMelf,and MelfandRazztogethercauseZlim. WouldZlimoccurifnotGlentinsteadofGlent? H:GlentcausesRazz,RazzandGlenttogethercausePex,PexcausesZurn,ZurncausesMelf,and MelfandRazztogethercauseZlim. WouldZlimoccurifnotGlentinsteadofGlent? I:GlentcausesRazz,RazzandGlenttogethercausePex,PexcausesZurn,ZurncausesMelf,and MelfandRazztogethercauseZlim. WouldZlimoccurifnotGlentinsteadofGlent? Figure13: ConclusionErrorExample Response Thecorrectansweris(0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0, 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0, Figure14: TypeMismatch