Causal-Audit: Explicit and Auditable Graph-based Reasoning via Target-Aware Causal Chain Construction

# Causal-Audit: Explicit and Auditable Graph-based Reasoning via Target-Aware Causal Chain Construction
Source: [https://arxiv.org/html/2607.15281](https://arxiv.org/html/2607.15281)
Alan Wee\-Chung Liew School of Information and Communication Technology, Griffith University, QLD, AU

###### Abstract

Causal and intervention\-based question answering is fundamental to advancing large language models \(LLMs\) toward reasoning beyond surface\-level correlations and understanding underlying causal mechanisms\. However, existing LLM\-based methods often rely on implicit language\-level reasoning, resulting in opaque causal assumptions, unverifiable reasoning paths, and fragile predictions under complex interventions, particularly in context\-free settings\. In this paper, we propose an explicit and auditable causal reasoning framework for context\-free intervention\-based question answering\. Our method formulates causal inference as structured reasoning over an explicit causal graph through four modular stages, rather than implicit end\-to\-end prediction\. A key innovation is a target\-aware causal graph construction strategy that treats the target variable as a core constraint during graph expansion, effectively suppressing irrelevant variables, spurious causal relations, and reasoning noise\. We further introduce a path\-level causal evidence aggregation mechanism that combines multiple causal paths while modeling both reinforcing and counteracting effects, enabling robust decision\-making beyond single\-chain reasoning\. Extensive experiments on three benchmarks demonstrate that our framework consistently outperforms existing LLM\-based methods while providing interpretable and auditable causal reasoning traces\.

Causal\-Audit: Explicit and Auditable Graph\-based Reasoning via Target\-Aware Causal Chain Construction

Su Lan, Xuefei Yin, Yanming Zhu, and Alan Wee\-Chung LiewSchool of Information and Communication Technology, Griffith University, QLD, AU

## 1Introduction

Causal and intervention\-based question answering is widely regarded as a critical step in advancing large language models \(LLMs\) from powerful language processors toward reliable reasoning systemsWuet al\.\([2024](https://arxiv.org/html/2607.15281#bib.bib44)\)\. Such questions ask how outcomes would change under hypothetical interventions rather than predicting what is typically observed, requiring models to reason about causal effects instead of exploiting statistical regularitiesJinet al\.\([2023](https://arxiv.org/html/2607.15281#bib.bib17)\); Zhouet al\.\([2024](https://arxiv.org/html/2607.15281#bib.bib21)\)\. Formally, it concerns the effect of actively manipulating a variable on a target outcome, often involving multiple interacting factors and indirect causal pathways\.

Among different settings,*context\-free*causal question answering, where models must infer causal effects without access to supporting passages or explicit background context, is particularly challengingPearl and Mackenzie \([2018](https://arxiv.org/html/2607.15281#bib.bib45)\)\. In this setting, answers cannot be derived via evidence extraction or surface\-level alignment, but instead require the internal construction and evaluation of plausible causal mechanisms\. In this work, we focus on this challenging yet fundamental setting of*context\-free causal question answering*\.

Recent advances in LLMs have led to encouraging performance gains on causal and counterfactual benchmarks, largely driven by end\-to\-end prediction or Chain\-of\-Thought style promptingWeiet al\.\([2022b](https://arxiv.org/html/2607.15281#bib.bib23)\); Kojimaet al\.\([2022](https://arxiv.org/html/2607.15281#bib.bib24)\)\. However, a growing body of empirical studies suggests that such improvements often arise from pattern matching or shallow heuristics rather than genuine causal understandingMilleret al\.\([2025](https://arxiv.org/html/2607.15281#bib.bib14)\)\. LLMs have been shown to struggle with intervention\-level reasoning, multi\-variable causal dependencies, and generalization beyond observed correlations, especially in settings that require reasoning about unseen or counterfactual scenariosJinet al\.\([2023](https://arxiv.org/html/2607.15281#bib.bib17)\); Wang \([2024](https://arxiv.org/html/2607.15281#bib.bib19)\); Zhouet al\.\([2024](https://arxiv.org/html/2607.15281#bib.bib21)\)\.

Several methods have been proposed to enhance LLMs for causal question answering, but most perform reasoning primarily at the language level, implicitly delegating causal inference to token\- or semantic\-level generationWeiet al\.\([2022b](https://arxiv.org/html/2607.15281#bib.bib23)\)\. Paradigms such as Chain\-of\-Thought \(CoT\)Weiet al\.\([2022a](https://arxiv.org/html/2607.15281#bib.bib42)\), Tree\-of\-ThoughtYaoet al\.\([2023](https://arxiv.org/html/2607.15281#bib.bib43)\), and Graph\-of\-ThoughtBestaet al\.\([2024b](https://arxiv.org/html/2607.15281#bib.bib26)\)organize textual rationales without explicit variable\-level causal structure or intervention semantics, making them vulnerable to spurious yet semantically related information\. Although some methods introduce causal supervision during trainingLiet al\.\([2025](https://arxiv.org/html/2607.15281#bib.bib9)\), inference typically collapses to direct question answering without explicit construction or auditing of causal graphs, leaving causal pathways, effect propagation, and interference from spurious variables unobservable and unverifiableGendronet al\.\([2024](https://arxiv.org/html/2607.15281#bib.bib46)\)\. Recent studies have consequently raised concerns about the reliability and faithfulness of such implicit reasoning, showing that generated rationales may not causally support final predictionsPaulet al\.\([2024](https://arxiv.org/html/2607.15281#bib.bib30)\); Chuet al\.\([2025](https://arxiv.org/html/2607.15281#bib.bib15)\)\. Moreover, methods that explicitly construct graphs from textual evidence rely on context\-rich inputs and often fail in context\-free settings, where relevant variables and relations cannot be reliably groundedTandonet al\.\([2019](https://arxiv.org/html/2607.15281#bib.bib11)\)\.

To address these limitations, we propose an explicit and auditable causal reasoning framework for intervention\-based question answering\. Rather than relying on implicit end\-to\-end prediction or free\-form rationales, we formulate causal reasoning as structured inference over an explicit causal graph, exposing intermediate causal hypotheses and enabling inspection of the reasoning process via four modular stages\. A key contribution is a*target\-aware*causal graph construction strategy that treats the target variable as a core constraint during graph expansion, effectively suppressing irrelevant variables, spurious relations, and reasoning noise\. Moreover, we introduce a path\-level causal evidence aggregation mechanism that combines multiple causal paths and models both reinforcing and counteracting effects, moving beyond single\-chain reasoning\. These designs transform LLMs from implicit end\-to\-end “reasoners" into constrained “causal evaluators", offering a new paradigm that integrates LLMs with symbolic causal reasoning for reliable intervention\-based question answering\. Extensive experiments on three datasets demonstrate consistent performance gains over existing LLM\-based methods while providing interpretable and auditable reasoning traces\.

The main contributions are summarized below:

- •We formulate context\-free intervention\-based question answering as an explicit and auditable causal reasoning problem, moving beyond implicit language\-level inference and providing a structured framework for reliable intervention reasoning with LLMs\.
- •We propose a target\-aware causal graph construction that explicitly treats the target variable as a core constraint during graph expansion, effectively suppressing irrelevant variables, spurious causal relations, and reasoning noise in traditional graph\-based reasoning\.
- •We introduce a path\-level causal evidence aggregation mechanism that validates and combines multiple causal paths while modeling both reinforcing and counteracting effects, enabling robust decision making beyond single\-chain or end\-to\-end methods\.
- •Extensive experiments on three benchmarks demonstrate that our framework consistently outperforms existing LLM\-based reasoning methods while providing interpretable and auditable causal reasoning traces\.

## 2Related Work

### 2\.1Direct LLM

Recent benchmarks such as CausalBenchWang \([2024](https://arxiv.org/html/2607.15281#bib.bib19)\); Zhouet al\.\([2024](https://arxiv.org/html/2607.15281#bib.bib21)\), CLadderJinet al\.\([2023](https://arxiv.org/html/2607.15281#bib.bib17)\), and CausalProbeChiet al\.\([2024](https://arxiv.org/html/2607.15281#bib.bib20)\)show that, despite strong surface\-level performance, LLMs often rely on shallow correlations rather than genuine causal understanding, especially under complex interventions\. Counterfactual and interventional reasoning further emerges inconsistently through in\-context learning and remains fragile to prompt design and distribution shiftsMilleret al\.\([2025](https://arxiv.org/html/2607.15281#bib.bib14)\)\. These reveal a persistent gap between correlation\-based inference and causal structure learning\.

![Refer to caption](https://arxiv.org/html/2607.15281v1/x1.png)Figure 1:Overview of the proposed explicit causal reasoning framework for context\-free intervention\-based question answering\. The framework extracts structured causal variables from the question, constructs a target\-aware causal graph via iterative search, audits candidate causal chains through counterfactual validation, and aggregates evidence from multiple stages to produce an interpretable intervention decision\. Detailed algorithmic descriptions are provided in Appendix[A](https://arxiv.org/html/2607.15281#A1)\.
### 2\.2Chain\-of\-Thought Reasoning

CoT prompting improves multi\-step reasoningWeiet al\.\([2022b](https://arxiv.org/html/2607.15281#bib.bib23)\); Kojimaet al\.\([2022](https://arxiv.org/html/2607.15281#bib.bib24)\), with self\-consistency further enhancing robustnessWanget al\.\([2023](https://arxiv.org/html/2607.15281#bib.bib25)\)\. However, CoT\-based methods rely on unstructured natural language rationales, lack causal and structural constraints, and are often not causally faithful to final predictionsPaulet al\.\([2024](https://arxiv.org/html/2607.15281#bib.bib30)\); Chuet al\.\([2025](https://arxiv.org/html/2607.15281#bib.bib15)\)\. Graph\-based prompting and neuro\-symbolic methods introduce structural scaffoldingBestaet al\.\([2024b](https://arxiv.org/html/2607.15281#bib.bib26)\); Fanget al\.\([2024](https://arxiv.org/html/2607.15281#bib.bib27)\)but do not explicitly encode causal semantics\. In contrast, our work grounds graph construction in causal relations and validates reasoning at the path level\.

### 2\.3Causal Graph\-based Reasoning

Integrating causal graphs and causal inference into LLM reasoning has emerged as a promising direction for improving faithfulness and interpretability\. Existing methods incorporate causal structuresWanget al\.\([2024](https://arxiv.org/html/2607.15281#bib.bib28)\)or employ causal objectives and structured supervisionDonget al\.\([2025](https://arxiv.org/html/2607.15281#bib.bib13)\); Liet al\.\([2025](https://arxiv.org/html/2607.15281#bib.bib9)\), yet intervention\-based analyses reveal frequent violations of causal faithfulnessPaulet al\.\([2024](https://arxiv.org/html/2607.15281#bib.bib30)\)\. Moreover, most methods rely on predefined or externally provided graphs\. In contrast, our framework dynamically constructs target\-aware causal graphs and performs path\-level validation and evidence aggregation during inference\.

### 2\.4Knowledge Graph\-Augmented Reasoning

Another line of work augments LLM reasoning with external knowledge graphs \(KGs\)\. Path\-based methods select informative KG paths to support multi\-hop reasoningLiuet al\.\([2024](https://arxiv.org/html/2607.15281#bib.bib31)\); Tanet al\.\([2025](https://arxiv.org/html/2607.15281#bib.bib33)\)\. Recent benchmarks evaluate LLMs’ use of KG structureMarkowitzet al\.\([2025](https://arxiv.org/html/2607.15281#bib.bib34)\), with causal\-aware graph augmentation further exploredLuoet al\.\([2025](https://arxiv.org/html/2607.15281#bib.bib16)\)\. However, these methods treat paths primarily as evidence and do not explicitly model causal direction or competing effects\. Our framework instead models signed causal influence and aggregates conflicting evidence during inference\.

## 3Methodology

Overview\.Given a context\-free intervention\-based causal question, the goal is to determine the directional effect of an intervention variable on a target outcome\. To this end, we propose an explicit causal reasoning framework that constructs, audits, and aggregates causal evidence in a structured and interpretable manner\. As illustrated in Fig\.[1](https://arxiv.org/html/2607.15281#S2.F1), the framework consists of four sequential stages\. First, structured query extraction identifies the intervention variable, target variables, and outcome modifiers from the input question\. Second, a target\-aware causal graph is constructed through iterative expansion, where graph growth is explicitly constrained by relevance to the target variable\. Third, candidate causal chains connecting the intervention and target are extracted from the graph and audited to ensure causal consistency under hypothetical interventions\. Finally, validated causal evidence from the first three stages is aggregated to produce the final decision on the direction of the intervention effect\. This staged formulation enables systematic control of the reasoning process, improves interpretability through explicit causal structures, and allows only causally validated evidence to contribute to the final prediction\.

### 3\.1Causal Variable Identification

Given a natural language questionqq, we employ an LLM\-based extractorΦextract\\Phi\_\{\\text\{extract\}\}to produce a structured query

𝒬=\(X,Ys,Yb,I\)\.\\mathcal\{Q\}=\(X,Y\_\{s\},Y\_\{b\},I\)\.\(1\)Here,XXdenotes the*intervention variable*representing the manipulated cause event;YsY\_\{s\}is the*surface target*, i\.e\., the surface\-level outcome expression in the question;YbY\_\{b\}is the*base target variable*, a canonical scientific noun phrase used for causal reasoning; andIIcontains*outcome modifiers*, including the directional indicatorD∈\{More,Less\}D\\in\\\{\\textsc\{More\},\\textsc\{Less\}\\\}and a negation flag\. We distinguishYsY\_\{s\}fromYbY\_\{b\}so that graph construction and causal reasoning are performed on the normalized variable pair\(X,Yb\)\(X,Y\_\{b\}\), while the final prediction is mapped back to the original question semantics usingII\. This separation decouples causal reasoning from surface linguistic variation, enabling more stable graph construction and consistent reasoning across semantically equivalent questions\.

### 3\.2Target\-Aware Causal Graph Construction

We propose to construct a directed causal graph𝒢=\(𝒱,ℰ\)\\mathcal\{G\}=\(\\mathcal\{V\},\\mathcal\{E\}\)rooted at the intervention variableXX, with the objective of identifying causal mechanisms leading to the base target variableYbY\_\{b\}\.

#### BFS Expansion with Negative Constraints\.

We adopt a breadth\-first search \(BFS\) expansion strategyKozen \([1992](https://arxiv.org/html/2607.15281#bib.bib47)\)that iteratively explores candidate causal effects layer by layer to construct the graph\. This process incrementally grows the graph outward fromXX, allowing the framework to capture multi\-hop causal mechanisms while maintaining explicit control over the search depth and expansion scope\. We initialize the frontierℱ0=\{X\}\\mathcal\{F\}\_\{0\}=\\\{X\\\}and the visited set𝒱vis=\{X\}\\mathcal\{V\}\_\{\\text\{vis\}\}=\\\{X\\\}, where the frontier represents the set of nodes to be expanded at the current step, and the visited set records all nodes observed so far\. At each expansion steptt, we collect a pool of LLM\-proposed candidate causal triples for the next layer, denoted by𝒞t\+1\\mathcal\{C\}\_\{t\+1\}:

𝒞t\+1\\displaystyle\\mathcal\{C\}\_\{t\+1\}=⋃u∈ℱt\{\(u,r,v\)\|\(u,r,v\)∼\\displaystyle=\\bigcup\_\{u\\in\\mathcal\{F\}\_\{t\}\}\\Bigl\\\{\(u,r,v\)\\ \\Bigm\|\\ \(u,r,v\)\\sim\(2\)Φexpand\(u∣Avoid=𝒱vis\)\}\.\\displaystyle\\qquad\\Phi\_\{\\text\{expand\}\}\\\!\\left\(u\\mid\\text\{Avoid\}=\\mathcal\{V\}\_\{\\text\{vis\}\}\\right\)\\Bigr\\\}\.Here,\(u,r,v\)\(u,r,v\)represents a directed causal relationu→vu\\\!\\rightarrow\\\!vwith effect polarityr∈\{Inc,Dec\}r\\in\\\{\\textsc\{Inc\},\\textsc\{Dec\}\\\}, whereIncdenotesu↑⇒v↑u\\uparrow\\Rightarrow v\\uparrowandDecdenotesu↑⇒v↓u\\uparrow\\Rightarrow v\\downarrow\. The∼\\simdenotes LLM generation or decoding\. Injecting the visited set as a negative constraint discourages revisiting previously explored nodes, reducing redundant loops and cyclic reasoning\. This design enforces acyclic graph growth and prevents uncontrolled expansion, keeping the search focused on plausible causal mechanisms\.

#### Fine\-Grained Variable Alignment \(FGVA\)\.

To determine whether a generated nodevvis relevant to the base targetYbY\_\{b\}, we introduce a fine\-grained variable alignment functionΦalign​\(v,Yb\)\\Phi\_\{\\text\{align\}\}\(v,Y\_\{b\}\)that evaluates three dimensions: \(i\)*Entity alignment*SE∈\{Exact,Partial,None\}S\_\{E\}\\in\\\{\\text\{Exact\},\\text\{Partial\},\\text\{None\}\\\}; \(ii\)*Quantity alignment*SQ∈\{Exact,Subset,Agg,None\}S\_\{Q\}\\in\\\{\\text\{Exact\},\\text\{Subset\},\\text\{Agg\},\\text\{None\}\\\}; and \(iii\)*State alignment*SS∈\{Match,Conflict,None\}S\_\{S\}\\in\\\{\\text\{Match\},\\text\{Conflict\},\\text\{None\}\\\}\. Intuitively,SES\_\{E\}assesses whether the core entity or topic matches \(or partially overlaps\),SQS\_\{Q\}checks whethervvandYbY\_\{b\}refer to the same granularity \(e\.g\., exact vs\. subset/aggregate\), andSSS\_\{S\}checks whether the described state is logically compatible\. By jointly considering these dimensions, FGVA enables relevance assessment beyond surface string similarity, allowing the framework to distinguish target\-equivalent variables, near\-target proxies, and intermediate bridging concepts\.

Based on the alignment results, we classify each nodevvinto a relevance class:

C​l​s​\(v\)=\{Exact,SEe​x∧SQe​x∧SSm,CloseHit,SEe​x∧SQs​a∧¬SSc,Bridge,SEp​a∧¬SSc,None,otherwise\.Cls\(v\)=\\begin\{cases\}\\textsc\{Exact\},&S\_\{E\}^\{ex\}\\land S\_\{Q\}^\{ex\}\\land S\_\{S\}^\{m\},\\\\ \\textsc\{CloseHit\},&S\_\{E\}^\{ex\}\\land S\_\{Q\}^\{sa\}\\land\\neg S\_\{S\}^\{c\},\\\\ \\textsc\{Bridge\},&S\_\{E\}^\{pa\}\\land\\neg S\_\{S\}^\{c\},\\\\ \\textsc\{None\},&\\text\{otherwise\}\.\\end\{cases\}

\(3\)whereSEe​x≜\[SE=Exact\]S\_\{E\}^\{ex\}\\\!\\triangleq\\\!\[S\_\{E\}=\\text\{Exact\}\],SEp​a≜\[SE=Partial\]S\_\{E\}^\{pa\}\\\!\\triangleq\\\!\[S\_\{E\}=\\text\{Partial\}\],SQe​x≜\[SQ=Exact\]S\_\{Q\}^\{ex\}\\\!\\triangleq\\\!\[S\_\{Q\}=\\text\{Exact\}\],SQs​a≜\[SQ∈\{Subset,Agg\}\]S\_\{Q\}^\{sa\}\\\!\\triangleq\\\!\[S\_\{Q\}\\in\\\{\\text\{Subset\},\\text\{Agg\}\\\}\],SSm≜\[SS=Match\]S\_\{S\}^\{m\}\\\!\\triangleq\\\!\[S\_\{S\}=\\text\{Match\}\], andSSc≜\[SS=Conflict\]S\_\{S\}^\{c\}\\\!\\triangleq\\\!\[S\_\{S\}=\\text\{Conflict\}\]\. The assigned classC​l​s​\(v\)Cls\(v\)determines how the node is used during graph construction\. Consequently, we maintain three auxiliary sets:𝒩exact=\{v∣C​l​s​\(v\)=Exact\}\\mathcal\{N\}\_\{\\text\{exact\}\}=\\\{v\\mid Cls\(v\)=\\textsc\{Exact\}\\\},𝒩close=\{v∣C​l​s​\(v\)=CloseHit\}\\mathcal\{N\}\_\{\\text\{close\}\}=\\\{v\\mid Cls\(v\)=\\textsc\{CloseHit\}\\\}, and𝒩bridge=\{v∣C​l​s​\(v\)=Bridge\}\\mathcal\{N\}\_\{\\text\{bridge\}\}=\\\{v\\mid Cls\(v\)=\\textsc\{Bridge\}\\\}, while nodes withC​l​s​\(v\)=NoneCls\(v\)=\\textsc\{None\}are discarded and not expanded further\.Exactnodes are treated as target\-equivalent and serve as terminal endpoints for causal\-chain extraction\.CloseHitnodes act as semantic landing points nearYbY\_\{b\}and support subsequent alignment toward the target\.Bridgenodes are retained only as intermediate candidates and may continue to be expanded to reachCloseHitorExactnodes, helping reduce semantic drift while preserving recall\.

#### Target\-Aware Pruning and Frontier Update\.

To control the exponential growth of the search space during graph expansion, we introduce a target\-aware pruning strategy at each BFS step\. Unconstrained expansion rapidly introduces loosely related or generic nodes, diluting causal signals and increasing the risk of spurious reasoning pathsSuiet al\.\([2022](https://arxiv.org/html/2607.15281#bib.bib40)\)\. Instead of heuristic string matching or static embedding similarity, we propose to use an LLM\-based relevance ranking functionΦrank​\(v,Yb\)\\Phi\_\{\\text\{rank\}\}\(v,Y\_\{b\}\)that explicitly conditions on the base target variableYbY\_\{b\}\. This ranking prioritizes nodes that are not only semantically related, but also causally promising for reaching the target domain\.

At each expansion step, only the top\-KKranked nodes are retained as the next frontier:

ℱt\+1=Top\-​K​\(𝒞t\+1\),𝒱v​i​s←𝒱v​i​s∪ℱt\+1\.\\mathcal\{F\}\_\{t\+1\}=\\text\{Top\-\}K\(\\mathcal\{C\}\_\{t\+1\}\),~~\\mathcal\{V\}\_\{vis\}\\leftarrow\\mathcal\{V\}\_\{vis\}\\cup\\mathcal\{F\}\_\{t\+1\}\.\(4\)This goal\-directed pruning focuses the expansion on plausible causal mechanisms while maintaining coverage of relevant effects\. The process terminates whenYbY\_\{b\}is reached or when a maximum depthDDis exceeded\.

#### Bridging Close\-Hit Nodes\.

When the BFS expansion does not explicitly reachYbY\_\{b\}, we perform a bridge step to connect near\-target nodes in𝒩c​l​o​s​e\\mathcal\{N\}\_\{close\}\. For eachu∈𝒩c​l​o​s​eu\\in\\mathcal\{N\}\_\{close\}, a verifierΨbridge\\Psi\_\{\\text\{bridge\}\}evaluates whether a direct causal relation is plausible; if so, a directed edge\(u,rbridge,Yb\)\(u,r\_\{\\text\{bridge\}\},Y\_\{b\}\)is added toℰ\\mathcal\{E\}\.

### 3\.3Causal Chain Extraction and Auditing

To derive causally valid explanations for intervention effects, we extract and audit causal chains connecting the intervention variableXXto the base targetYbY\_\{b\}from the constructed graph𝒢\\mathcal\{G\}\. We enumerate simple paths𝒫=\{p1,…,pm\}\\mathcal\{P\}=\\\{p\_\{1\},\\dots,p\_\{m\}\\\}connectingXXandYbY\_\{b\}, rank them using a structural cost function that penalizes excessive length, semantic drift, and generic nodes, and retain the top\-K′K^\{\\prime\}paths for a two\-stage causal audit\.

#### Premise Consistency Check\.

A valid causal explanation must not contradict the intervention premise\. For a pathpp, we define a premise validity indicator

Vprem​\(p\)=∏v∈p∖\{X,Yb\}𝕀​\[¬C​\(v,X\)\]\.V\_\{\\text\{prem\}\}\(p\)=\\prod\_\{v\\in p\\setminus\\\{X,Y\_\{b\}\\\}\}\\mathbb\{I\}\\\!\\left\[\\neg C\(v,X\)\\right\]\.\(5\)whereC​\(v,X\)C\(v,X\)indicates if introducing an intermediate nodevvwould violate the intervention premise onXX\. Paths withVprem​\(p\)=0V\_\{\\text\{prem\}\}\(p\)=0are discarded\.

#### Counterfactual Edge Verification\.

Each remaining path is further audited via counterfactual probing\. For each edgee=\(u,v\)e=\(u,v\), an LLM estimates whether removing or reversinguuinduces a change invv, producing a edge\-level confidenceVcf​\(e\)∈\[0,1\]V\_\{\\text\{cf\}\}\(e\)\\in\[0,1\]\. The final path\-level audit score is defined as

Saudit​\(p\)=Vprem​\(p\)⋅∏e∈pVcf​\(e\)\.S\_\{\\text\{audit\}\}\(p\)=V\_\{\\text\{prem\}\}\(p\)\\cdot\\prod\_\{e\\in p\}V\_\{\\text\{cf\}\}\(e\)\.\(6\)Paths with multiplicative audit scoreSaudit​\(p\)S\_\{\\text\{audit\}\}\(p\)less than a predefined thresholdτaudit\\tau\_\{\\text\{audit\}\}are removed\.

### 3\.4Evidence Aggregation and Final Decision

#### Path\-Level Effect and Weighting\.

Each causal relationrris assigned a signed effectσ​\(r\)∈\{\+1,−1\}\\sigma\(r\)\\in\\\{\+1,\-1\\\}\. The net causal effect of a pathppis computed as

Δ​\(p\)=∏e∈pσ​\(re\),\\Delta\(p\)=\\prod\_\{e\\in p\}\\sigma\(r\_\{e\}\),\(7\)reflecting whether the intervention increases or decreases the target outcome\. To downweight weakly supported or implicit reasoning, each path is assigned a weight

w​\(p\)=Saudit​\(p\)⋅exp⁡\(−γ​Nbr​\(p\)\),w\(p\)=S\_\{\\text\{audit\}\}\(p\)\\cdot\\exp\(\-\\gamma N\_\{\\text\{br\}\}\(p\)\),\(8\)whereNbr​\(p\)N\_\{\\text\{br\}\}\(p\)denotes the number of bridge edges\.

#### Graph\-Level Evidence Aggregation\.

We aggregate signed evidence across all audited paths𝒫∗\\mathcal\{P\}^\{\*\}\. LetW\+W^\{\+\}andW−W^\{\-\}denote the total positive and negative evidence mass, respectively\. The resulting graph\-level confidence is defined as

Cgraph=\|W\+−W−\|W\+\+W−\+ϵ⋅log⁡\(1\+\|𝒫∗\|\),C\_\{\\text\{graph\}\}=\\frac\{\|W^\{\+\}\-W^\{\-\}\|\}\{W^\{\+\}\+W^\{\-\}\+\\epsilon\}\\cdot\\log\\\!\\left\(1\+\|\\mathcal\{P\}^\{\*\}\|\\right\),\(9\)which jointly captures evidence dominance and support size\. The corresponding graph decisionEgraph∈\{More,Less\}E\_\{\\text\{graph\}\}\\in\\\{\\textsc\{More\},\\textsc\{Less\}\\\}is determined by the sign ofW\+−W−W^\{\+\}\-W^\{\-\}\.

#### Conflict Resolution with Evidence\-Conditioned LLM\.

When causal evidence is weak or conflicting, we defer to an LLM prediction conditioned on the audited graph evidence\. We serialize𝒫∗\\mathcal\{P\}^\{\*\}into a compact evidence packet𝒵​\(𝒫∗\)\\mathcal\{Z\}\(\\mathcal\{P\}^\{\*\}\), containing for each path its signed effectΔ​\(p\)\\Delta\(p\), weightw​\(p\)w\(p\), and a short edge\-sequence representation\.111Deterministic decoding is used\.Formally,

ELLM=fθ​\(q,𝒵​\(𝒫∗\)\)∈\{More,Less\}\.E\_\{\\text\{LLM\}\}=f\_\{\\theta\}\\\!\\big\(q,\\mathcal\{Z\}\(\\mathcal\{P\}^\{\*\}\)\\big\)\\in\\\{\\textsc\{More\},\\textsc\{Less\}\\\}\.\(10\)The final prediction is obtained by resolvingEgraphE\_\{\\text\{graph\}\}andELLME\_\{\\text\{LLM\}\}using the graph confidence:

y^=\{Egraph,Cgraph\>τ​or​Egraph=ELLM,ELLM,otherwise\.\\hat\{y\}=\\begin\{cases\}E\_\{\\text\{graph\}\},&C\_\{\\text\{graph\}\}\>\\tau\\ \\text\{or\}\\ E\_\{\\text\{graph\}\}=E\_\{\\text\{LLM\}\},\\\\ E\_\{\\text\{LLM\}\},&\\text\{otherwise\}\.\\end\{cases\}

\(11\)Finally,y^\\hat\{y\}is mapped to the surface\-level answer using the modifier setII\.

## 4Experiments

### 4\.1Datasets, Task, and Evaluation Metric

Datasets and Task\.We evaluate our method on three benchmarks spanning diverse domains and knowledge requirements: \(i\) DDXPlus\-CausalEffect \(medical\); \(ii\) the directional subset of WIQA; and \(iii\) a general\-domain dataset derived from CauseNet\. All datasets are cast into a unified*context\-free causal direction prediction*setting\. Each instance specifies an intervention on a cause variable or eventXXand asks whether the target variable or eventYYbecomesmoreorlesslikely under the intervention\. Following prior protocols on WIQA, we exclude NoEff instances and retain the original two\-choice answer space\. For DDXPlus\-CausalEffect and the CauseNet\-derived dataset, questions are rendered using a WIQA\-style template with the same label set\{more,less\}\\\{\\texttt\{more\},\\texttt\{less\}\\\}\. This unification ensures consistent inference and fair comparison across datasets and methods\. Dataset construction details and statistics are provided in Appendix[B](https://arxiv.org/html/2607.15281#A2)\.

Evaluation Metric\.Following prior works, we report*Accuracy*as the evaluation metric\. Since the task is a binary directional classification with a fixed label space and follows standard evaluation protocols on the directional subset, Accuracy provides a clear and comparable measure across backbones and methods\. Label distributions are reported in Appendix[B](https://arxiv.org/html/2607.15281#A2)for completeness\.

### 4\.2Experimental Setup

Decoding and Reproducibility\.All LLM calls use deterministic decoding \(temperature=0=0\) with fixed decoding hyperparameters \(e\.g\.,top\_pandmax\_tokens\)\. When supported by the backend, we additionally fix the random seed to improve run\-to\-run consistency\. Outputs are strictly constrained via JSON/schema validation; invalid responses are retried using a predefined template fallback\.

Prompt Format\.All methods adopt structured prompts with a fixed output schema, where the final prediction is constrained to\{more,less\}\\\{\\texttt\{more\},\\texttt\{less\}\\\}and optional fields are used for intermediate reasoning or evidence\. Full prompts, schemas, and example instances are provided in Appendix[C](https://arxiv.org/html/2607.15281#A3)\.

Graph Construction and Auditing\.Graph expansion is performed with bounded depthD=4D=4and Top\-K=2K=2frontier selection per hop\. For each frontier node, up toR=3R=3candidate relations are generated, and simple paths fromXXto the \(base\) target are enumerated up to lengthL=6L=6\. Candidate paths are filtered using an audit thresholdτaudit=0\.6\\tau\_\{\\text\{audit\}\}=0\.6before evidence aggregation\. All hyperparameters are selected on a held\-out development split \(or via the grid search in Sec\.[4\.5](https://arxiv.org/html/2607.15281#S4.SS5)\) and fixed across all test runs\.

### 4\.3Backbone Models and Compared Methods

Backbone ModelsWe evaluate our method on multiple instruction\-tuned LLM backbones from different model families at comparable scales, including Llama\-3\.1\-8B, Mistral\-7B, and Ministral\-3\-8B\. To isolate the effect of the proposed reasoning framework, we keep the dataset, instance representation \(explicit\(X,Y\)\(X,Y\)\), prompt format, input/output schema, and decoding hyperparameters fixed for each backbone, and vary only the backbone model\. This allows us to assess whether performance gains generalize across model families rather than relying on a specific backbone\.

Compared MethodsWe compare our method with representative LLM\-based reasoning baselines under a unified evaluation pipeline\. Specifically, we include \(i\)*Direct LLM*, which directly predicts the directional label from\(X,Y\)\(X,Y\)without intermediate structure, serving as a parametric\-knowledge baseline; \(ii\)*Chain\-of\-Thought*\(COT\) promptingWeiet al\.\([2022a](https://arxiv.org/html/2607.15281#bib.bib42)\), which elicits free\-form step\-by\-step rationales before prediction; \(iii\) structured reasoning methods such as*Tree\-of\-Thought*\(ToT\)Yaoet al\.\([2023](https://arxiv.org/html/2607.15281#bib.bib43)\)and*Graph\-of\-Thought*\(GoT\)Bestaet al\.\([2024a](https://arxiv.org/html/2607.15281#bib.bib39)\), which explore multiple reasoning paths via tree\- or graph\-structured deliberation without explicit causal modeling or verification; and \(iv\) CDCR\-SFTLiet al\.\([2025](https://arxiv.org/html/2607.15281#bib.bib9)\), which enhances causal reasoning by supervised LoRA fine\-tuning on CausalDR, encouraging LLMs to explicitly construct causal DAGs before making predictions\. All baselines use the same final answer constraint\{more,less\}\\\{\\texttt\{more\},\\texttt\{less\}\\\}and identical decoding settings unless otherwise specified\.

Table 1:Accuracy \(%\) on context\-free causal direction prediction across three datasets and multiple LLM backbones\. Free\-form reasoning methods \(CoT, GoT, ToT\) exhibit high variance across backbones\. Ours consistently achieves the best performance across backbones and domains, demonstrating robust gains from explicit causal graph construction, auditing, and evidence aggregation\. Accuracy is reported under a best\-of\-multiple\-attempts setting\.Table 2:Accuracy \(%\) on WIQA \(direction subset\) compared with CDCR\-SFT, a training\-based method that learns causal DAG construction via supervised fine\-tuning\. Our method achieves substantially higher accuracy on both backbones without task\-specific fine\-tuning, highlighting the effectiveness of explicit and auditable causal modeling at inference time\.
### 4\.4Main Results

Table[1](https://arxiv.org/html/2607.15281#S4.T1)reports accuracy \(%\) for context\-free causal direction prediction across datasets and backbones\. Overall,oursconsistently outperforms all baselines\. We highlight three recurring patterns\.

\(1\) Reliability under domain\-specific causal knowledge\.On DDXPlus\-CausalEffect, which involves specialized medical knowledge and non\-trivial causal dependencies, unconstrained reasoning baselines \(CoT, ToT, GoT\) are often unstable and can underperform Direct LLM prediction\. This suggests that free\-form rationales tend to introduce unsupported or domain\-invalid mechanisms in expert domains\. In contrast,oursconsistently improves performance across all backbones, with particularly large gains on Mistral\-7B \(\+28\.5\) and Ministral \(\+21\.0\)\. These results indicate that our proposed explicit construction and auditing of causal chains is crucial for reliable intervention reasoning in knowledge\-intensive settings\.

\(2\) Strong gains on complex multi\-hop causal reasoning\.On WIQA, which requires multi\-hop causal reasoning under interventions,oursyields substantial improvements over the strongest baseline across all backbones \(\+11\.8, \+23\.6, and \+4\.3 points for Llama\-3\.1\-8B, Mistral\-7B, and Ministral, respectively\)\. Notably, existing reasoning paradigms exhibit high variance across backbones, whereasoursconsistently achieves the best performance\. This suggests that our proposed explicit construction and auditing of causal structures is more effective than relying on implicit deliberation alone, particularly for long\-horizon causal inference\. By converting global reasoning into target\-aware expansion and local edge verification, our framework provides a reliable inductive bias that generalizes across model families\.

\(3\) Robustness in open\-domain causal reasoning\.On the open\-domain CauseNet\-derived benchmark, where spurious co\-occurrence is common, purely deliberative or search\-based reasoning methods exhibit high variance across backbones\. For example, ToT performs competitively on Llama\-3\.1\-8B but collapses on Mistral\-7B\. In contrast,oursconsistently achieves strong performance across all backbones, outperforming the best baseline on Mistral\-7B \(\+6\.0\) and Ministral \(\+8\.0\)\. This robustness highlights the benefit of our proposed target\-aware pruning and evidence aggregation in suppressing spurious causal paths and maintaining reliable inference in noisy, open\-domain settings\.

Comparison with causal\-structure fine\-tuning\.To further contextualize our WIQA results, we compare against CDCR\-SFT, a recent method that improves causal reasoning by supervised LoRA fine\-tuning on the CausalDR dataset to explicitly learn causal DAG construction during training\. As shown in Table[2](https://arxiv.org/html/2607.15281#S4.T2),ourssubstantially outperforms CDCR\-SFT on both backbones, with gains of \+12\.26 on Llama\-3\.1\-8B and \+20\.28 on Mistral\-7B\. Notably, our approach achieves these improvements without any task\-specific fine\-tuning, demonstrating that explicit and auditable causal modeling at inference time can be more effective and generalizable than embedding causal structure implicitly into model parameters\.

Table 3:Hyperparameter sensitivity analysis on WIQA \(direction subset\) with the Llama\-3\.1\-8B backbone\.
### 4\.5Hyperparameter Sensitivity Analysis

We analyze the sensitivity of our target\-aware causal graph construction to three key hyperparameters: maximum expansion depth \(D∈\{2,4,6\}D\\in\\\{2,4,6\\\}\), branching width \(R∈\{3,5\}R\\in\\\{3,5\\\}\), and maximum path length \(L∈\{4,6\}L\\in\\\{4,6\\\}\)\. Results are reported in Table[3](https://arxiv.org/html/2607.15281#S4.T3)onWIQAwith the Llama\-3\.1\-8B backbone\.

Effect of Expansion Depth\.Performance improves when increasing the depth fromD=2D=2to a moderate range \(D=4D=4\), reflecting the necessity of multi\-hop causal mechanisms for WIQA\-style intervention questions\. However, further increasing depth \(D=6D=6\) yields diminishing or unstable gains, suggesting that overly deep expansion introduces semantically weak or noisy intermediate variables\. This trend supports our design choice of bounded expansion with auditing, which prioritizes causal relevance over exhaustive search\.

Interaction Between Branching and Path Length\.We observe a clear trade\-off between branching width \(RR\) and allowed path length \(LL\)\. Moderate branching with sufficient path length \(e\.g\.,R=3,L=6R=3,L=6\) consistently yields strong performance, suggesting that accurate intervention reasoning relies on precise multi\-hop causal chains rather than broad, shallow exploration\. In contrast, aggressive early branching often degrades accuracy, particularly for shorter paths, due to the introduction of spurious or weakly related variables\. These trends directly motivate our target\-aware pruning and path\-level auditing: by constraining early expansion and validating longer chains, our framework favors causally coherent multi\-hop mechanisms while suppressing noisy alternatives\.

Exogenous vs\. Internal Interventions\.Exo\-Acc and In\-Acc measure accuracy when the intervention variable appears outside or inside the paragraph, respectively\. Higher In\-Acc reflects better alignment and reduced semantic drift, while gains on Exo\-Acc indicate stronger context\-free generalization\. Exo\-Acc benefits from larger expansion depth due to longer mechanism chains, whereas In\-Acc typically peaks at moderate depth\. Their complementary depth preferences validate our bounded, target\-aware expansion design\.

### 4\.6Case Study and Interpretability

We present a representative case study to illustrate how explicit causal modeling improves both reliability and interpretability\. While standard CoT reasoning produces free\-form narratives whose causal validity is difficult to assess, our framework externalizes reasoning as explicit causal graphs, audited chains, and quantified evidence\. In a DDXPlus\-CausalEffect example predicting the effect onChagas probability, CoT incorrectly outputsmoreby invoking a generic prior \(“early\-stage disease may be asymptomatic”\) that is not causally grounded in the intervention\. In contrast, our method constructs multiple candidate causal chains from the intervention variable, filters invalid mechanisms via counterfactual edge verification, and aggregates only premise\-consistent evidence\. This process yields the correctlessprediction and exposes the specific causal paths supporting the decision\. We provide the full constructed graph, audited chains, and per\-edge counterfactual scores in Appendix[D](https://arxiv.org/html/2607.15281#A4)\.

## 5Conclusion

We propose an explicit, auditable causal reasoning framework for context\-free intervention question answering\. It formulates inference as structured reasoning via target\-aware causal graph construction, path\-level auditing, and evidence aggregation\. By constraining variable expansion and verifying causal relations via counterfactual probing, it transparently resolves competing causal mechanisms\. This design offers a clean paradigm for integrating LLMs with symbolic causal reasoning, transforming LLMs from end\-to\-end reasoners into constrained causal evaluators\. Extensive experiments across medical, commonsense, and open\-domain benchmarks demonstrate consistent and substantial improvements over existing methods, underscoring the importance of explicit causal structure and verification for reliable causal reasoning\.

## Limitations

Inference ComplexityThe framework relies on multiple LLM interactions to support explicit causal graph construction and verification\. Compared to single\-pass approaches, this introduces additional inference steps, but the process is bounded by controlled expansion depth and pruning, reflecting a deliberate trade\-off between reasoning reliability and efficiency\.

Potential Risk of Misuse\.The framework produces directional causal predictions with structured reasoning traces, but does not perform formal causal identification based on do\-calculus\. In sensitive domains such as healthcare, the outputs should therefore be interpreted as supportive reasoning signals rather than definitive causal or clinical conclusions\.

## References

- M\. Besta, N\. Blach, A\. Kubicek, R\. Gerstenberger, M\. Podstawski, L\. Gianinazzi, J\. Gajda, T\. Lehmann, H\. Niewiadomski, P\. Nyczyk,et al\.\(2024a\)Graph of thoughts: solving elaborate problems with large language models\.InProceedings of the AAAI conference on artificial intelligence,Vol\.38,pp\. 17682–17690\.Cited by:[§4\.3](https://arxiv.org/html/2607.15281#S4.SS3.p2.2)\.
- M\. Besta, N\. Blach, A\. Kubicek, R\. Gerstenberger, M\. Podstawski, L\. Gianinazzi, J\. Gajda, T\. Lehmann, H\. Niewiadomski, P\. Nyczyk,et al\.\(2024b\)Graph of thoughts: solving elaborate problems with large language models\.InProceedings of the AAAI conference on artificial intelligence,Vol\.38,pp\. 17682–17690\.Cited by:[§1](https://arxiv.org/html/2607.15281#S1.p4.1),[§2\.2](https://arxiv.org/html/2607.15281#S2.SS2.p1.1)\.
- A\. Bondarenko, M\. Wolska, S\. Heindorf, L\. Blübaum, A\. N\. Ngomo, B\. Stein, P\. Braslavski, M\. Hagen, and M\. Potthast \(2022\)CausalQA: a benchmark for causal question answering\.InProceedings of the 29th International Conference on Computational Linguistics,pp\. 3296–3308\.Cited by:[§B\.1](https://arxiv.org/html/2607.15281#A2.SS1.p1.3)\.
- H\. Chi, H\. Li, W\. Yang, F\. Liu, L\. Lan, X\. Ren, T\. Liu, and B\. Han \(2024\)Unveiling causal reasoning in large language models: reality or mirage?\.Vol\.37,pp\. 96640–96670\.Cited by:[§2\.1](https://arxiv.org/html/2607.15281#S2.SS1.p1.1)\.
- Z\. Chu, J\. Chen, Z\. Wang, G\. Tang, Q\. Chen, M\. Liu, and B\. Qin \(2025\)Towards faithful multi\-step reasoning through fine\-grained causal\-aware attribution reasoning distillation\.InProceedings of the 31st International Conference on Computational Linguistics,pp\. 2291–2315\.Cited by:[§1](https://arxiv.org/html/2607.15281#S1.p4.1),[§2\.2](https://arxiv.org/html/2607.15281#S2.SS2.p1.1)\.
- J\. Dong, Y\. Liu, A\. Aloui, V\. Tarokh, and D\. Carlson \(2025\)CARE: turning llms into causal reasoning expert\.InFirst Workshop on Foundations of Reasoning in Language Models,Cited by:[§2\.3](https://arxiv.org/html/2607.15281#S2.SS3.p1.1)\.
- M\. Fang, S\. Deng, Y\. Zhang, Z\. Shi, L\. Chen, M\. Pechenizkiy, and J\. Wang \(2024\)Large language models are neurosymbolic reasoners\.InProceedings of the AAAI conference on artificial intelligence,Vol\.38,pp\. 17985–17993\.Cited by:[§2\.2](https://arxiv.org/html/2607.15281#S2.SS2.p1.1)\.
- A\. Fansi Tchango, R\. Goel, Z\. Wen, J\. Martel, and J\. Ghosn \(2022\)Ddxplus: a new dataset for automatic medical diagnosis\.Advances in neural information processing systems35,pp\. 31306–31318\.Cited by:[§B\.2](https://arxiv.org/html/2607.15281#A2.SS2.p1.4)\.
- G\. Gendron, J\. M\. Rožanec, M\. Witbrock, and G\. Dobbie \(2024\)Counterfactual causal inference in natural language with large language models\.arXiv preprint arXiv:2410\.06392\.Cited by:[§1](https://arxiv.org/html/2607.15281#S1.p4.1)\.
- S\. Heindorf, Y\. Scholten, H\. Wachsmuth, A\. Ngonga Ngomo, and M\. Potthast \(2020\)Causenet: towards a causality graph extracted from the web\.InProceedings of the 29th ACM international conference on information & knowledge management,pp\. 3023–3030\.Cited by:[§B\.4](https://arxiv.org/html/2607.15281#A2.SS4.p1.3)\.
- Z\. Jin, Y\. Chen, F\. Leeb, L\. Gresele, O\. Kamal, Z\. Lyu, K\. Blin, F\. Gonzalez Adauto, M\. Kleiman\-Weiner, M\. Sachan,et al\.\(2023\)Cladder: assessing causal reasoning in language models\.Advances in Neural Information Processing Systems36,pp\. 31038–31065\.Cited by:[§B\.1](https://arxiv.org/html/2607.15281#A2.SS1.p1.3),[§1](https://arxiv.org/html/2607.15281#S1.p1.1),[§1](https://arxiv.org/html/2607.15281#S1.p3.1),[§2\.1](https://arxiv.org/html/2607.15281#S2.SS1.p1.1)\.
- T\. Kojima, S\. S\. Gu, M\. Reid, Y\. Matsuo, and Y\. Iwasawa \(2022\)Large language models are zero\-shot reasoners\.Advances in neural information processing systems35,pp\. 22199–22213\.Cited by:[§1](https://arxiv.org/html/2607.15281#S1.p3.1),[§2\.2](https://arxiv.org/html/2607.15281#S2.SS2.p1.1)\.
- D\. C\. Kozen \(1992\)Depth\-first and breadth\-first search\.InThe design and analysis of algorithms,pp\. 19–24\.Cited by:[§3\.2](https://arxiv.org/html/2607.15281#S3.SS2.SSS0.Px1.p1.5)\.
- Y\. Li, Y\. Shen, Y\. Nian, J\. Gao, Z\. Wang, C\. Yu, S\. Li, J\. Wang, X\. Hu, and Y\. Zhao \(2025\)Mitigating hallucinations in large language models via causal reasoning\.InFindings of the Association for Computational Linguistics: ACL 2025,pp\. 16563–16577\.Cited by:[§1](https://arxiv.org/html/2607.15281#S1.p4.1),[§2\.3](https://arxiv.org/html/2607.15281#S2.SS3.p1.1),[§4\.3](https://arxiv.org/html/2607.15281#S4.SS3.p2.2)\.
- H\. Liu, S\. Wang, Y\. Zhu, Y\. Dong, and J\. Li \(2024\)Knowledge graph\-enhanced large language models via path selection\.InFindings of the Association for Computational Linguistics ACL 2024,pp\. 6311–6321\.Cited by:[§2\.4](https://arxiv.org/html/2607.15281#S2.SS4.p1.1)\.
- H\. Luo, J\. Zhang, and C\. Li \(2025\)Causal graphs meet thoughts: enhancing complex reasoning in graph\-augmented llms\.arXiv preprint arXiv:2501\.14892\.Cited by:[§2\.4](https://arxiv.org/html/2607.15281#S2.SS4.p1.1)\.
- E\. Markowitz, K\. Galiya, G\. V\. Steeg, and A\. Galstyan \(2025\)KG\-llm\-bench: a scalable benchmark for evaluating llm reasoning on textualized knowledge graphs\.arXiv preprint arXiv:2504\.07087\.Cited by:[§2\.4](https://arxiv.org/html/2607.15281#S2.SS4.p1.1)\.
- M\. Miller, B\. Schölkopf, and S\. Guo \(2025\)Counterfactual reasoning: an analysis of in\-context emergence\.InThirty\-Ninth Annual Conference on Neural Information Processing Systems,Cited by:[§1](https://arxiv.org/html/2607.15281#S1.p3.1),[§2\.1](https://arxiv.org/html/2607.15281#S2.SS1.p1.1)\.
- D\. Paul, R\. West, A\. Bosselut, and B\. Faltings \(2024\)Making reasoning matter: measuring and improving faithfulness of chain\-of\-thought reasoning\.InFindings of the Association for Computational Linguistics: EMNLP 2024,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),pp\. 15012–15032\.Cited by:[§1](https://arxiv.org/html/2607.15281#S1.p4.1),[§2\.2](https://arxiv.org/html/2607.15281#S2.SS2.p1.1),[§2\.3](https://arxiv.org/html/2607.15281#S2.SS3.p1.1)\.
- J\. Pearl and D\. Mackenzie \(2018\)The book of why: the new science of cause and effect\.Basic Books\.Cited by:[§1](https://arxiv.org/html/2607.15281#S1.p2.1)\.
- Y\. Sui, S\. Feng, H\. Zhang, J\. Cao, L\. Hu, and N\. Zhu \(2022\)Causality\-aware enhanced model for multi\-hop question answering over knowledge graphs\.Knowledge\-Based Systems250,pp\. 108943\.External Links:ISSN 0950\-7051Cited by:[§3\.2](https://arxiv.org/html/2607.15281#S3.SS2.SSS0.Px3.p1.2)\.
- X\. Tan, X\. Wang, Q\. Liu, X\. Xu, X\. Yuan, and W\. Zhang \(2025\)Paths\-over\-graph: knowledge graph empowered large language model reasoning\.InProceedings of the ACM on Web Conference 2025,pp\. 3505–3522\.Cited by:[§2\.4](https://arxiv.org/html/2607.15281#S2.SS4.p1.1)\.
- N\. Tandon, B\. Dalvi, K\. Sakaguchi, P\. Clark, and A\. Bosselut \(2019\)WIQA: a dataset for “what if…” reasoning over procedural text\.InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing \(EMNLP\-IJCNLP\),pp\. 6076–6085\.Cited by:[§B\.3](https://arxiv.org/html/2607.15281#A2.SS3.p1.1),[§1](https://arxiv.org/html/2607.15281#S1.p4.1)\.
- J\. Wang, D\. Cao, S\. Lu, Z\. Ma, J\. Xiao, and T\. Chua \(2024\)Causal\-driven large language models with faithful reasoning for knowledge question answering\.InProceedings of the 32nd ACM International Conference on Multimedia,pp\. 4331–4340\.Cited by:[§2\.3](https://arxiv.org/html/2607.15281#S2.SS3.p1.1)\.
- X\. Wang, J\. Wei, D\. Schuurmans, Q\. V\. Le, E\. H\. Chi, S\. Narang, A\. Chowdhery, and D\. Zhou \(2023\)Self\-consistency improves chain of thought reasoning in language models\.Cited by:[§2\.2](https://arxiv.org/html/2607.15281#S2.SS2.p1.1)\.
- Z\. Wang \(2024\)CausalBench: a comprehensive benchmark for evaluating causal reasoning capabilities of large language models\.InProceedings of the 10th SIGHAN Workshop on Chinese Language Processing,pp\. 143–151\.Cited by:[§1](https://arxiv.org/html/2607.15281#S1.p3.1),[§2\.1](https://arxiv.org/html/2607.15281#S2.SS1.p1.1)\.
- J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, F\. Xia, E\. Chi, Q\. V\. Le, D\. Zhou,et al\.\(2022a\)Chain\-of\-thought prompting elicits reasoning in large language models\.Advances in neural information processing systems35,pp\. 24824–24837\.Cited by:[§1](https://arxiv.org/html/2607.15281#S1.p4.1),[§4\.3](https://arxiv.org/html/2607.15281#S4.SS3.p2.2)\.
- J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, F\. Xia, E\. Chi, Q\. V\. Le, D\. Zhou,et al\.\(2022b\)Chain\-of\-thought prompting elicits reasoning in large language models\.Advances in neural information processing systems35,pp\. 24824–24837\.Cited by:[§1](https://arxiv.org/html/2607.15281#S1.p3.1),[§1](https://arxiv.org/html/2607.15281#S1.p4.1),[§2\.2](https://arxiv.org/html/2607.15281#S2.SS2.p1.1)\.
- A\. Wu, K\. Kuang, M\. Zhu, Y\. Wang, Y\. Zheng, K\. Han, B\. Li, G\. Chen, F\. Wu, and K\. Zhang \(2024\)Causality for large language models\.arXiv preprint arXiv:2410\.15319\.Cited by:[§1](https://arxiv.org/html/2607.15281#S1.p1.1)\.
- S\. Yao, D\. Yu, J\. Zhao, I\. Shafran, T\. Griffiths, Y\. Cao, and K\. Narasimhan \(2023\)Tree of thoughts: deliberate problem solving with large language models\.Advances in neural information processing systems36,pp\. 11809–11822\.Cited by:[§1](https://arxiv.org/html/2607.15281#S1.p4.1),[§4\.3](https://arxiv.org/html/2607.15281#S4.SS3.p2.2)\.
- Y\. Zhou, X\. Wu, B\. Huang, J\. Wu, L\. Feng, and K\. C\. Tan \(2024\)Causalbench: a comprehensive benchmark for causal learning capability of llms\.arXiv preprint arXiv:2404\.06349\.Cited by:[§1](https://arxiv.org/html/2607.15281#S1.p1.1),[§1](https://arxiv.org/html/2607.15281#S1.p3.1),[§2\.1](https://arxiv.org/html/2607.15281#S2.SS1.p1.1)\.

Appendix

## Appendix AAlgorithm Pseudocode

Algorithm[1](https://arxiv.org/html/2607.15281#algorithm1)presents the pseudocode of the proposed framework\.

Input :Question

qq
Output :Predicted answer

y^\\hat\{y\}
//1\. Setup \(Sec\.[3\.1](https://arxiv.org/html/2607.15281#S3.SS1)\)

\(X,Ys,Yb,I\)←Φextract​\(q\)\(X,Y\_\{s\},Y\_\{b\},I\)\\leftarrow\\Phi\_\{\\text\{extract\}\}\(q\);

ℱ←\{X\}\\mathcal\{F\}\\leftarrow\\\{X\\\};

𝒱v​i​s←\{X\}\\mathcal\{V\}\_\{vis\}\\leftarrow\\\{X\\\};

𝒩c​l​o​s​e←∅\\mathcal\{N\}\_\{close\}\\leftarrow\\emptyset;

//2\. Graph Expansion \(Sec\.[3\.2](https://arxiv.org/html/2607.15281#S3.SS2)\)

while*Yb∉ℱY\_\{b\}\\notin\\mathcal\{F\}anddepth<D<D*do

//Generate candidates avoiding history

𝒞←Φexpand​\(ℱ∣Avoid=𝒱v​i​s\)\\mathcal\{C\}\\leftarrow\\Phi\_\{\\text\{expand\}\}\(\\mathcal\{F\}\\mid\\text\{Avoid\}=\\mathcal\{V\}\_\{vis\}\);

//Identify Close Hits via FGVA

foreach*v∈𝒞v\\in\\mathcal\{C\}*do

if*Φ*align*​\(v,Yb\)=*CloseHit*\\Phi\_\{\\text\{align\}\}\(v,Y\_\{b\}\)=\\textsc\{CloseHit\}*then

𝒩c​l​o​s​e←𝒩c​l​o​s​e∪\{v\}\\mathcal\{N\}\_\{close\}\\leftarrow\\mathcal\{N\}\_\{close\}\\cup\\\{v\\\};

//Prune to Top\-K

ℱ←Φrank​\(𝒞,Yb,K\)\\mathcal\{F\}\\leftarrow\\Phi\_\{\\text\{rank\}\}\(\\mathcal\{C\},Y\_\{b\},K\);

𝒱v​i​s←𝒱v​i​s∪ℱ\\mathcal\{V\}\_\{vis\}\\leftarrow\\mathcal\{V\}\_\{vis\}\\cup\\mathcal\{F\};

Update Graph

𝒢\\mathcal\{G\};

//3\. Bridging \(Sec\.[3\.2](https://arxiv.org/html/2607.15281#S3.SS2)\)

foreach*u∈𝒩c​l​o​s​eu\\in\\mathcal\{N\}\_\{close\}*do

if*Ψ*bridge*​\(u,Yb\)\\Psi\_\{\\text\{bridge\}\}\(u,Y\_\{b\}\)is valid*then

Add edge

\(u→Yb\)\(u\\to Y\_\{b\}\)to

𝒢\\mathcal\{G\};

//4\. Extraction & Auditing \(Sec\.[3\.3](https://arxiv.org/html/2607.15281#S3.SS3)\)

𝒫∗←Top\-kPaths​\(𝒢,X→Yb\)\\mathcal\{P\}^\{\*\}\\leftarrow\\text\{Top\-kPaths\}\(\\mathcal\{G\},X\\to Y\_\{b\}\);

foreach*p∈𝒫∗p\\in\\mathcal\{P\}^\{\*\}*do

//Check Premise & Counterfactuals

if*V*prem*​\(p\)=0V\_\{\\text\{prem\}\}\(p\)=0orV*cf*​\(p\)<τV\_\{\\text\{cf\}\}\(p\)<\\tau*then

Discard path

pp;

//5\. Decision \(Sec\.[3\.4](https://arxiv.org/html/2607.15281#S3.SS4)\)

Egraph←AggregatePaths​\(𝒫∗\)E\_\{\\text\{graph\}\}\\leftarrow\\text\{AggregatePaths\}\(\\mathcal\{P\}^\{\*\}\);

ELLM←Φreason​\(q,𝒫∗\)E\_\{\\text\{LLM\}\}\\leftarrow\\Phi\_\{\\text\{reason\}\}\(q,\\mathcal\{P\}^\{\*\}\);

y^←ResolveConflict​\(Egraph,ELLM\)\\hat\{y\}\\leftarrow\\text\{ResolveConflict\}\(E\_\{\\text\{graph\}\},E\_\{\\text\{LLM\}\}\);

return

y^\\hat\{y\}mapped to options via

II;

Algorithm 1Target\-Aware Causal Graph Reasoning
## Appendix BDataset Construction Details

### B\.1Overview

Most existing causal question answering resources are*context\-dependent*, where each question is paired with a passage and models are evaluated on evidence extractionJinet al\.\([2023](https://arxiv.org/html/2607.15281#bib.bib17)\); Bondarenkoet al\.\([2022](https://arxiv.org/html/2607.15281#bib.bib18)\)\. In contrast, we target*context\-free causal direction prediction*with*graph\-grounded explanations*: each instance is defined over explicit variables/events\(X,Y\)\(X,Y\)and the system predicts whether intervening onXXmakesYY*more*or*less*likely\.

We focus on the directional subset with labels\{more,less\}\\\{\\texttt\{more\},\\texttt\{less\}\\\}\. Some datasets includeno effectas a distractor option, but we excludeNoEffinstances as gold labels to keep the evaluation protocol consistent\. Table[4](https://arxiv.org/html/2607.15281#A2.T4)summarizes dataset statistics reports how often questions include*surface modifiers*\(e\.g\., “more/less probability ofYY”\)\.

Table 4:Dataset statistics\.Directional subset only \(more/less\)\. Negation via keyword match \(e\.g\.,*not/no/without*\) on the question stem\.
### B\.2DDXPlus\-CausalEffect: Directional Effect Prediction from Medical Records

We use the official DDXPlus releaseFansi Tchangoet al\.\([2022](https://arxiv.org/html/2607.15281#bib.bib10)\), including condition metadata, evidence metadata, and patient\-level predefined splits\. Each instance specifies an intervention variableXX\(a patient evidence\) and an outcome variableYY\(a pathology\), and asks how changingXXaffects the probability ofYY\.

#### Association\-based labeling \(Stats Mode\)\.

Our primary construction assigns labels using empirical association statistics computed from patient records\. For each pathologyYYand evidenceXX, we computenYn\_\{Y\},nXn\_\{X\},nX​Yn\_\{XY\}, andNN\(total patients\), and estimate:

p​\(Y∣X\)=nX​YnX,p​\(Y∣¬X\)=nY−nX​YN−nX,p\(Y\\mid X\)=\\frac\{n\_\{XY\}\}\{n\_\{X\}\},\\qquad p\(Y\\mid\\neg X\)=\\frac\{n\_\{Y\}\-n\_\{XY\}\}\{N\-n\_\{X\}\},\(12\)then defineΔ=p​\(Y∣X\)−p​\(Y∣¬X\)\\Delta=p\(Y\\mid X\)\-p\(Y\\mid\\neg X\)\. Using a margin thresholdτ\\tau, we assign:

label​\(X,Y\)=\{more,Δ\>τless,Δ<−τdiscard,\|Δ\|≤τ\.\\texttt\{label\}\(X,Y\)=\\begin\{cases\}\\texttt\{more\},&\\Delta\>\\tau\\\\ \\texttt\{less\},&\\Delta<\-\\tau\\\\ \\texttt\{discard\},&\|\\Delta\|\\leq\\tau\.\\end\{cases\}\(13\)These labels are treated asdata\-driven directional proxies\(association\-based\) rather than fully identified causal effects; discarding near\-zero cases avoids ambiguous instances when statistics are insufficient\.

#### Question rendering\.

We render each pair\(X,Y\)\(X,Y\)into a WIQA\-style question with answer choices restricted to\{more,less\}\\\{\\texttt\{more\},\\texttt\{less\}\\\}\(optionally keepingno effectonly as a distractor\)\. An optional LLM rewriting step can improve fluency while preserving explicitXXandYY\.

### B\.3WIQA Directional Subset

WIQATandonet al\.\([2019](https://arxiv.org/html/2607.15281#bib.bib11)\)is the closest widely\-used benchmark aligned with our evaluation format because it is framed as effect\-direction prediction and its questions are formulated over explicit standalone events\. Following prior work, we evaluate on a curated subset that contains only directional\-effect instances \(more/less\) and excludesNoEffcases\.

### B\.4CauseNet\-derived Context\-free Directional QA

We construct a general\-domain dataset from CauseNetHeindorfet al\.\([2020](https://arxiv.org/html/2607.15281#bib.bib12)\)\. Each record provides a head–tail causal assertion with optional confidence\-like fields\. We \(i\) sanitize concepts, \(ii\) normalize heterogeneous confidence signals to a unified edge confidencec​\(e\)∈\[0,1\]c\(e\)\\in\[0,1\], and \(iii\) map relations to signed directed edges withsign∈\{\+1,−1\}\\mathrm\{sign\}\\in\\\{\+1,\-1\\\}\(defaulting to\+1\+1if polarity is missing\)\.

#### Multi\-hop path sampling and labeling\.

We sample simple directed pathsπ=\(x0→⋯→xh\)\\pi=\(x\_\{0\}\\\!\\rightarrow\\\!\\cdots\\\!\\rightarrow\\\!x\_\{h\}\)and derive the gold direction label by polarity product:

label​\(π\)=\{more,∏e∈πsign​\(e\)=\+1,less,∏e∈πsign​\(e\)=−1\.\\texttt\{label\}\(\\pi\)=\\begin\{cases\}\\texttt\{more\},&\\prod\_\{e\\in\\pi\}\\mathrm\{sign\}\(e\)=\+1,\\\\ \\texttt\{less\},&\\prod\_\{e\\in\\pi\}\\mathrm\{sign\}\(e\)=\-1\.\\end\{cases\}\(14\)We keep only directional instances to align with our evaluation protocol\.

#### Question rendering\.

For each sampled path, we form a WIQA\-style context\-free question over endpointsX=x0X=x\_\{0\}andY=xhY=x\_\{h\}and store the underlying sampled path as an auditable explanation trace\.

#### Limitations\.

This construction inherits noise from open IE extraction and may underrepresent negative relations when explicit inhibitory edges are sparse; we therefore report label distributions and key sampling hyperparameters alongside results\.

## Appendix CPrompt Templates and Output Constraints

### C\.1Common I/O Format

All datasets are cast into a unifiedcontext\-free directional\-effectformat: each instance specifies explicit variables/events\(X,Y\)\(X,Y\)and asks whether intervening onXXmakesYYmoreorlesslikely\. Following prior WIQA protocols, we excludeNoEffand keep atwo\-choicesetting\. Accordingly, all prompts in this paper share the same answer space:A: more,B: less\. All LLM calls use deterministic decoding withtemperature=0\.

### C\.2Answer Extraction and Fallback

For all methods, we enforce an explicit final answer line:Final answer: AorFinal answer: B\. We extract the choice using a regex matcher\. If the required line is missing, we apply a lightweight*forced extractor*prompt that maps the model output to\{A,B\}\\\{A,B\\\}; if extraction still fails, we fall back toA\.

### C\.3Baseline Prompts

#### Direct LLM\.

The model receives only the question stem and the two options, and must output the final choice:

> \[Direct\] \{question\_stem\} A\) more B\) less Output format: Final answer: <A\|B\>

#### Chain\-of\-Thought \(CoT\)\.

We elicit a brief rationale before producing the final choice, while keeping the same answer constraint:

> \[CoT\] Write brief causal reasoning, then choose A/B\. Output format: Reasoning: <1\-\-4 sentences\> Final answer: <A\|B\>

#### Graph\-of\-Thought \(GoT\)\.

GoT follows a split→\\rightarrowanalyze→\\rightarrowmerge procedure with explicit intermediate artifacts\. We first request a JSON decomposition into three components; then we analyze each component and produce a merged solution\. All stages are deterministic and the final response must end with the forced choice:

> Split: Output ONLY JSON \{Component 1,2,3\}\. Analyze each component\. Merge and MUST end with: Final answer: <A\|B\>\.

#### Tree\-of\-Thought \(ToT\)\.

ToT performs a deterministic tree\-style deliberation with explicit candidate generation and selection: \(i\) generate two candidate reasoning structures; \(ii\) select the better one; \(iii\) generate two candidate reasoning traces conditioned on the selected structure; \(iv\) select the better trace; \(v\) output the final choice\. The final output is forced into\{A,B\}\\\{A,B\\\}:

> Generate 2 candidates \(JSON\)→\\rightarrowPick 1/2 Generate 2 traces \(JSON\)→\\rightarrowPick 1/2 Final: MUST end with: Final answer: <A\|B\>

We refer to these as GoT/ToT because they enforce graph\-/tree\-structured intermediate representations and selection steps, even though decoding is deterministic in our implementation\.

### C\.4Our Framework: Graph\-grounded \+ Audit Prompts

Our method uses the LLM as a*structured evidence worker*and a final*evidence aggregator*\. All intermediate prompts are constrained to JSON outputs to ensure robust parsing, while the final decision is mapped to the unified two\-choice space \(A/B\)\.

#### P1: Variable extraction \(strict JSON\)\.

```
You are a STRICT string-matching text
extractor.
Your job is ONLY to extract spans from
the given question text.
Do NOT paraphrase. Copy spans exactly.

Question:
"{question}"

Output ONLY valid JSON (no extra text):
{
  "cause_event": "<copied text>",
  "outcome_text_raw": "<copied text>",
  "outcome_direction": "MORE" or "LESS"
  or "NONE",
  "is_negated": true or false
}
```

#### P2: Target\-aware single\-hop expansion \(forward\)\.

```
You are a causal edge finder.

Input:
- CAUSE_NODE (X): "{X}"
- TARGET_HINT (Y): "{target_hint}"
- FORBIDDEN LIST (Avoid revisits):
[{avoid_str}]

Task:
- Propose up to {max_relations}
SINGLE-HOP causal effects
starting from X.
- If TARGET_HINT is not "NONE",
expand toward Y (prefer intermediates
that connect X to Y).
- Each tail must be a NEUTRAL NOUN
PHRASE (no "more/less", no full
sentences).
- Use "INCREASES" when increasing
head tends to increase tail.
- Use "DECREASES" when increasing
head tends to decrease tail.

Output ONLY JSON:
{
  "triples": [
    ["{X}", "INCREASES" | "DECREASES",
    "<neutral noun phrase>"]
  ]
}
```

#### P3: Target equivalence / bridging judgment\.

```
You are judging the relationship
between two variables in a causal system.

Variable A: "{A}"
Variable B: "{B}"

Task:
Decide their relationship along
three axes:
1) core_entity_relation
2) quantity_relation
3) causal_or_structural_relation

Output ONLY JSON:
{
  "core_entity_relation": "...",
  "quantity_relation": "...",
  "causal_or_structural_relation":
  "...",
  "explanation": "short explanation"
}
```

#### P4: Counterfactual edge audit \(binary validity\)\.

```
You are a Scientific Logic Judge.

We MUST judge causality based on
intervention semantics:
If we actively increase A,
does B tend to increase/decrease?

Candidate causal edge:
A: "{A}"
Relation: "{REL}"   (INCREASES/DECREASES)
B: "{B}"

Return a conservative judgment. If unsure,
mark false.

Output ONLY JSON:
{"is_valid_link": true/false, "reasoning":
"short explanation"}
```

#### P5: Final aggregation with audited chains\.

```
You are solving a WIQA-style causal
reasoning problem.
Your job is to decide how the CAUSE affects
the BASE VARIABLE.

Question: "{question}"
Cause event (X): "{cause_event}"
Outcome event (surface): "{outcome_event}"
BASE variable (outcome_base,
the only quantity you judge): "{outcome_base}"

Causal graph summary:
{summary_json}

Evidence chains from cause 
BASE (system-computed net effects;
DO NOT re-multiply signs yourself):
{evidence_block}

IMPORTANT:
- Decide the direction of change for
the BASE VARIABLE only ("{outcome_base}").
- [Net Effect: POSITIVE (Causes Increase)]
supports "more" (A).
- [Net Effect: NEGATIVE (Causes Decrease)]
supports "less" (B).
- If chains conflict, prefer higher-quality
/ fewer-bridge chains.

Output ONLY strict JSON:
{
  "effect_on_base": "more" | "less",
  "final_answer": "A" | "B",
  "confidence": "high" | "medium"
  | "low" | "very_low",
  "reasoning": "short explanation grounded in
  the evidence chains"
}
```

#### Example evidence\_block \(serialized chains\)\.

```
- Chain: X -> INCREASES -> M1 ->
DECREASES -> outcome_base
  [Net Effect: NEGATIVE (Causes Decrease)]
  [bridge_edges: 0/2]
- Chain: X -> INCREASES -> M2 -> INCREASES
-> outcome_base
  [Net Effect: POSITIVE (Causes Increase)]
  [bridge_edges: 1/2]
```

## Appendix DExtended Case Study

#### Instance\.

Intervention:X=X=the patient has not noticed any new fatigue, vague discomfort, diffuse muscle aches, or a change in well\-being\. Target:Yb=Y\_\{b\}=Chagas probability\. Gold label:less\.

#### Baseline \(CoT\) output\.

CoT predictsmoreby invoking a generic narrative that Chagas can be asymptomatic in early stages, treating symptom absence as evidence for infection\.

#### Representative generated causal triples\.

Our graph construction produces localized, inspectable hypotheses as directed triples\. Examples include:

- •X→DecX\\xrightarrow\{\\textsc\{Dec\}\}patient’s overall health status
- •X→IncX\\xrightarrow\{\\textsc\{Inc\}\}Chagas disease risk factors
- •Chagas disease risk factors→Inc\\xrightarrow\{\\textsc\{Inc\}\}infection probability
- •infection probability→Dec\\xrightarrow\{\\textsc\{Dec\}\}Chagas probability
- •Chagas disease risk factors→Inc\\xrightarrow\{\\textsc\{Inc\}\}Chagas probability
- •patient’s overall health status→Inc\\xrightarrow\{\\textsc\{Inc\}\}Chagas probability

#### Extracted causal chains\.

From the constructed graph, our path extractor returns three chains:

\(P1\)X→Decoverall health status→IncYb,\(P2\)X→Incrisk factors→IncYb,\(P3\)X→Incrisk factors→Incinfection probability→DecYb\.\\begin\{array\}\[\]\{ll\}\\text\{\(P1\)\}&X\\xrightarrow\{\\textsc\{Dec\}\}\\texttt\{overall health status\}\\xrightarrow\{\\textsc\{Inc\}\}Y\_\{b\},\\\\ \\text\{\(P2\)\}&X\\xrightarrow\{\\textsc\{Inc\}\}\\texttt\{risk factors\}\\xrightarrow\{\\textsc\{Inc\}\}Y\_\{b\},\\\\ \\text\{\(P3\)\}&X\\xrightarrow\{\\textsc\{Inc\}\}\\texttt\{risk factors\}\\\\ &\\xrightarrow\{\\textsc\{Inc\}\}\\texttt\{infection probability\}\\xrightarrow\{\\textsc\{Dec\}\}Y\_\{b\}\.\\end\{array\}

#### Counterfactual edge audit\.

For each edgee=\(u,v\)e=\(u,v\)on a candidate path, we compute a counterfactual support scoreVcf​\(e\)∈\[0,1\]V\_\{\\text\{cf\}\}\(e\)\\in\[0,1\]via intervention\-style probing, and define the path audit score as

Saudit​\(p\)=Vprem​\(p\)⋅∏e∈pVcf​\(e\)\.S\_\{\\text\{audit\}\}\(p\)=V\_\{\\text\{prem\}\}\(p\)\\cdot\\prod\_\{e\\in p\}V\_\{\\text\{cf\}\}\(e\)\.\(15\)In this example, the retained edges receive consistently high counterfactual support \(approximately0\.90\.9–1\.01\.0\), and the final decision is made by aggregating the audited paths\.

#### Aggregation and decision\.

The audited chains contain directional disagreement \(one positive vs\. two negative\)\. Our conflict\-aware aggregation weights each chain by its audit score and combines positive/negative evidence mass, resulting in a netlesseffect onYbY\_\{b\}and a final confidence of 0\.85, matching the gold label\.

## Appendix EEthics Statement

This manuscript is the authors’ original work\. Except for minor English grammar checking with ChatGPT, no large language model or AI tool was used for idea generation, problem formulation, literature search or screening, methodology design, code implementation, data processing, experimental design, statistical analysis, figure or table drafting, or substantive writing\. All intellectual contributions, including conceptualization, model design, and empirical evaluation, are solely those of the authors\.