DARE: Dialectical Agentic Reasoning for Structured Knowledge Fact Checking

# Dialectical Agentic Reasoning for Structured Knowledge Fact Checking
Source: [https://arxiv.org/html/2609.13808](https://arxiv.org/html/2609.13808)
\\tl\_set:Ne\\tcboxmath

tcboxmath\\tl\_set:Ne\\tcbhighmathtcbhighmath

Yifei LiXiaohan ZhengAffiliation:University of Science and Technology of ChinaEmail:[lszhuang@ustc\.edu\.cn](mailto:)Wentao QianAffiliation:University of Science and Technology of ChinaLiansheng Zhuang††thanks:Corresponding author\.Affiliation:University of Science and Technology of China

###### Abstract

Structured knowledge fact checking aims to determine the truthfulness of natural language claims by reasoning over structured evidence\. Recent program\-generation approaches leverage large language models \(LLMs\) to generate executable graph reasoning programs, achieving strong performance on structured knowledge fact checking benchmarks\. However, these methods remain limited by invalid relation generation, single\-path reasoning that lacks self\-correction, and biased evidence assessment that tends to overestimate supporting signals\. We proposeDialecticalAgenticREasoning \(DARE\), a multi\-agent framework that formulates structured knowledge fact checking as an iterative retrieve–reason–reflect process\. DARE integrates relation\-grounded evidence retrieval to constrain reasoning to valid structures, dialectical bidirectional verification to evaluate evidence from both supporting and refuting perspectives, and confidence\-driven meta\-reflection to dynamically determine whether additional evidence exploration is necessary\. Extensive experiments demonstrate the effectiveness of DARE in structured knowledge fact checking, with an 8B backbone achieving 88\.12% accuracy and matching or surpassing GPT\-4o\-based program\-generation baselines, which attests to the efficacy of dialectical agentic reasoning in eliciting the latent reasoning capabilities of LLMs\.

![Refer to caption](https://arxiv.org/html/2609.13808v1/DARE.png)Figure 1:Interaction among the retriever, reasoner, and controller agents\.## 1Introduction

Structured knowledge fact checking aims to determine whether a natural language claim is supported or refuted by structured knowledge in knowledge graphs\(kim2023factkg;thorne2018fever\)\. Given a claim, the system must retrieve relevant evidence from the knowledge graph and produce a grounded binary decision\. This task plays an important role in combating misinformation and has attracted increasing attention in the NLP community\(guo2022survey;nakov2021factcheck\)\.

Recent advances in large language models \(LLMs\) have shifted the paradigm from training specialized classifiers\(zhou2019gear\)to program\-generation approaches\(hao2025pgr;kim2023kggpt;pan2023programfc\), which prompt LLMs to generate executable reasoning programs composed of graph operations \(e\.g\.,SEARCH,MATCH,VERIFY\) and execute them over the knowledge graph to obtain verification results\.

Despite their success, we identify a fundamental limitation of existing program\-generation methods: they produce ungrounded and uncorrected reasoning trajectories over knowledge graphs\. In particular, reasoning failures arise from three tightly coupled issues\. First, LLMs are unaware of relation vocabularies specific to knowledge graphs, leading to invalid relation generation and execution errors\. Second, reasoning is performed in a single forward pass without mechanisms for iterative correction, making the process brittle under multi\-hop or compositional claims\. Third, verification is conducted from a single perspective, resulting in systematic overestimation of supporting evidence and poor confidence calibration\.

To address this limitation, we propose DARE \(Dialectical Agentic REasoning\), a unified dialectical reasoning framework that performs iterative, grounded verification over knowledge graphs\. Instead of treating fact verification as one\-shot program execution, DARE formulates it as a closed\-loop retrieve–reason–reflect process that couples evidence grounding with adversarial evaluation\. The interaction among the three agents is illustrated in Figure[1](https://arxiv.org/html/2609.13808#S0.F1)\.

Specifically, DARE first introduces relation\-grounded retrieval, which constrains reasoning to valid KG structures by retrieving candidate relations from the graph and filtering them according to claim semantics, ensuring that all reasoning steps are grounded in actual KG relations\. It then designs dialectical bidirectional reasoning, which evaluates evidence from both supporting and opposing perspectives to explicitly expose contradictions and reduce confirmation bias\. Finally, it proposes confidence\-driven meta\-reflection, which monitors the agreement between the two perspectives and adaptively determines whether additional reasoning iterations are required, forming a unified loop that continuously refines both evidence and judgment\.

- •We identify ungrounded and uncorrected reasoning trajectories as a key limitation of program\-generation methods for structured knowledge fact checking, and propose DARE, a dialectical reasoning framework that addresses this issue through grounded retrieval, dialectical evaluation, and adaptive reflection\.
- •We demonstrate that dialectical bidirectional reasoning improves evidence calibration, while confidence\-driven meta\-reflection enables iterative self\-correction during inference, leading to more reliable multi\-step reasoning\.
- •On the FactKG benchmark, DARE with a Qwen3\-8B model achieves 88\.12% accuracy, matching or exceeding GPT\-4o\-based program\-generation baselines, while using a substantially smaller backbone model\.

## 2Related Work

#### Fact Checking on Knowledge Graphs\.

Fact checking seeks to assess the truthfulness of claims based on structured or unstructured evidence\(thorne2018fever;guo2022survey\)\. Early benchmarks primarily rely on unstructured text\(thorne2018fever;aly2021feverous\)or semi\-structured tables\(chen2019tabfact\), leaving KG\-based verification underrepresented until FactKG\(kim2023factkg\), which grounds claims in the DBpedia knowledge graph\(lehmann2015dbpedia\)across multiple reasoning types\.

Traditional approaches train specialized classifiers over graph structures\(zhou2019gear\)\. More recently, LLM\-based methods have shifted toward few\-shot and program\-generation paradigms\. KG\-GPT\(kim2023kggpt\)and StructGPT\(jiang2023structgpt\)perform step\-by\-step reasoning via in\-context learning, ProgramFC\(pan2023programfc\)decomposes claims into verifiable sub\-claims, and PGR\(hao2025pgr\)defines graph reasoning primitives and prompts LLMs to compose executable programs\. A parallel line of work explores LLM–KG integration for broader reasoning tasks: Think\-on\-Graph\(sun2024tog\)iteratively explores KG paths through beam search, and RoG\(luo2024rog\)generates faithful reasoning chains grounded in KG relations\. Baek et al\.\(baek2024kalm\)augment LLMs with retrieved knowledge for verification but do not incorporate dialectical evaluation\. Despite their differences, these methods largely rely on a single forward reasoning trajectory over the KG, without explicit mechanisms for iterative correction or counterfactual evaluation\.

#### LLM Reasoning and Self\-Reflection\.

LLM reasoning has progressed from chain\-of\-thought prompting\(wei2022cot;kojima2022large\)to self\-consistency\(wang2023selfconsistency\), structured search over multiple reasoning paths\(yao2024tot;besta2024got\), and reasoning\-action integration\(yao2023react\)\. Self\-reflection further enables models to critique and refine their outputs through verbal feedback\(shinn2023reflexion\), self\-revision\(madaan2023selfrefine\), or tool\-based critique mechanisms\(gou2024critic\)\. Chain\-of\-Verification\(dhuliawala2023cove\)addresses hallucination by generating verification questions and checking consistency, though it operates over unstructured text rather than symbolic knowledge\. Multi\-agent systems extend this paradigm by enabling debate or role\-based collaboration to improve factuality\(du2023improving;liang2023encouraging;sun2024llmqa\)\.

However, these approaches predominantly rely on either self\-evaluation or consensus\-seeking dynamics, both of which remain vulnerable to miscalibrated confidence and self\-evaluation distortion\(kadavath2022calibration;tian2023justask\)\. Importantly, they lack structured counterfactual reasoning over external symbolic evidence such as knowledge graphs, which limits their effectiveness in verification tasks requiring grounded and adversarial evaluation\.

![Refer to caption](https://arxiv.org/html/2609.13808v1/DARE_WORKFLOW.png)Figure 2:Overview of the DARE framework for structured knowledge fact checking\. The retriever iteratively extracts relation\-grounded evidence from the knowledge graph\. The reasoner performs dialectical bidirectional reasoning over supporting and refuting perspectives to produce evidence\-aware assessments\. The controller applies confidence\-driven meta\-reflection to determine whether to terminate with a final prediction or continue iterative reasoning with expanded entities\.

## 3Method

### 3\.1Task Formulation

Given a knowledge graph𝒢=\(ℰ,ℛ,𝒯\)\\mathcal\{G\}=\(\\mathcal\{E\},\\mathcal\{R\},\\mathcal\{T\}\), whereℰ\\mathcal\{E\}denotes the entity set,ℛ\\mathcal\{R\}the relation set, and𝒯⊆ℰ×ℛ×ℰ\\mathcal\{T\}\\subseteq\\mathcal\{E\}\\times\\mathcal\{R\}\\times\\mathcal\{E\}the triple set, and a natural language claimssmentioning a subset of entitiesℰs⊆ℰ\\mathcal\{E\}\_\{s\}\\subseteq\\mathcal\{E\}, the goal of KG fact verification is to predict a label:

y=f⁡\(s,𝒢\)∈\{Sup,Ref\},y=f\(s,\\mathcal\{G\}\)\\in\\\{\\textsc\{Sup\},\\textsc\{Ref\}\\\},\(1\)whereSupandRefdenote*Supported*and*Refuted*, respectively\.

The task requires reasoning over the knowledge graph to determine whether the claim is supported by existing facts or contradicted by them\. Figure[2](https://arxiv.org/html/2609.13808#S2.F2)provides an overview of the proposed framework\.

### 3\.2Retrieval Module

The Retriever is responsible for gathering relevant evidence from the KG\. To avoid the relation mapping failures inherent in program\-generation approaches, the Retriever adopts a*retrieve\-then\-filter*strategy that operates in two stages\.

#### Adjacency Retrieval\.

For each entitye∈ℰse\\in\\mathcal\{E\}\_\{s\}mentioned in the claim, the Retriever retrieves all adjacent triples from𝒢\\mathcal\{G\}via direct neighborhood lookup:

Γe=\{\(h,r,o\)∈𝒯∣h=e∨o=e\},\\Gamma\_\{e\}=\\\{\(h,r,o\)\\in\\mathcal\{T\}\\mid h=e\\lor o=e\\\},\(2\)yielding the complete neighborhood subgraphΓs=⋃e∈ℰsΓe\\Gamma\_\{s\}=\\bigcup\_\{e\\in\\mathcal\{E\}\_\{s\}\}\\Gamma\_\{e\}\. This ensures that all retrieved relations exist in the KG, eliminating relation mapping errors\. At iterationt\>1t\>1, the Retriever expands the entity set to include entities discovered in previously retrieved triples, enabling multi\-hop evidence chains\.

#### Semantic Filtering\.

The full adjacency setΓs\\Gamma\_\{s\}often contains relations irrelevant to verifying the claim\. We employ an LLM\-based evidence selectorGEG\_\{E\}to semantically filter the most relevant triples based on the claim context:

E\(t\)=GE​\(Γs\(t\),s,θE\),E^\{\(t\)\}=G\_\{E\}\(\\Gamma\_\{s\}^\{\(t\)\},s;\\theta\_\{E\}\),\(3\)whereθE\\theta\_\{E\}denotes the role\-specific prompt that instructs the Retriever to select the top\-kkmost relevant triples fromΓs\(t\)\\Gamma\_\{s\}^\{\(t\)\}for verifyingss\. Specifically, the LLM ranks adjacent triples by relevance and selects top\-kkasE\(t\)⊆Γs\(t\)E^\{\(t\)\}\\subseteq\\Gamma\_\{s\}^\{\(t\)\}\. This retains only semantically relevant relations, reducing noise for subsequent reasoning while ensuring that all considered relations are valid KG entries\.

### 3\.3Dialectical Reasoning Module

The Reasoner performs*dialectical bidirectional reasoning*on the collected evidence to produce calibrated confidence assessments\. Rather than evaluating evidence from a single perspective—which is susceptible to the self\-evaluation distortion identified in Section 1—the Reasoner independently argues from both supporting and refuting viewpoints\.

#### Dialectical Bidirectional Reasoning\.

Given the accumulated evidenceE1:t=⋃i=1tE\(i\)E\_\{1:t\}=\\bigcup\_\{i=1\}^\{t\}E^\{\(i\)\}at iterationtt, the Reasoner employs an LLM\-based dialectical evaluatorDDto assess the evidence from two opposing perspectives\. Both branches take the same accumulated evidenceE1:tE\_\{1:t\}as input but are guided by different prompts\. First, the Reasoner generates a supporting argumentαt\+\\alpha\_\{t\}^\{\+\}and produces a supporting confidence score:

ct\+=D\+\(E1:t,s;θD\+\),c\_\{t\}^\{\+\}=D^\{\+\}\(E\_\{1:t\},s;\\theta\_\{D\}^\{\+\}\),\(4\)whereD\+D^\{\+\}denotes the supporting evaluation branch parameterized by promptθD\+\\theta\_\{D\}^\{\+\}, andct\+∈\[0,1\]c\_\{t\}^\{\+\}\\in\[0,1\]quantifies the degree to which the evidence supports the claim\. Independently, the Reasoner generates a refuting argumentαt−\\alpha\_\{t\}^\{\-\}and produces a refuting confidence score:

ct−=D−\(E1:t,s;θD−\),c\_\{t\}^\{\-\}=D^\{\-\}\(E\_\{1:t\},s;\\theta\_\{D\}^\{\-\}\),\(5\)whereD−D^\{\-\}denotes the refuting evaluation branch parameterized by promptθD−\\theta\_\{D\}^\{\-\}, andct−∈\[0,1\]c\_\{t\}^\{\-\}\\in\[0,1\]quantifies the degree to which the evidence refutes the claim\. The two promptsθD\+\\theta\_\{D\}^\{\+\}andθD−\\theta\_\{D\}^\{\-\}share the same structure but carry opposing argumentation directives:θD\+\\theta\_\{D\}^\{\+\}instructs the model to argue why the evidence supports the claim, whileθD−\\theta\_\{D\}^\{\-\}instructs it to argue why the evidence refutes the claim\.

By forcing the model to articulate both supporting and refuting arguments with equal rigor, dialectical reasoning counteracts self\-evaluation distortion: even when evidence has a superficial supporting tendency, the model must confront potential contradictions, producing more balanced and reliable confidence estimates\.

#### Cumulative Evidence Aggregation\.

At each iteration, the evidence and arguments from previous rounds are accumulated\. The cumulative supporting and refuting confidences at iterationttare computed as:

Ct\+=max1≤i≤t⁡ci\+,Ct−=max1≤i≤t⁡ci−\.C\_\{t\}^\{\+\}=\\max\_\{1\\leq i\\leq t\}c\_\{i\}^\{\+\},\\quad C\_\{t\}^\{\-\}=\\max\_\{1\\leq i\\leq t\}c\_\{i\}^\{\-\}\.\(6\)
The corresponding best supporting and refuting arguments are selected from the iterations that achieve these maxima:

αt\+⁣∗=αi\+\+,αt−⁣∗=αi−−,\\alpha\_\{t\}^\{\+\*\}=\\alpha\_\{i^\{\+\}\}^\{\+\},\\quad\\alpha\_\{t\}^\{\-\*\}=\\alpha\_\{i^\{\-\}\}^\{\-\},\(7\)wherei\+=arg⁡max1≤i≤t⁡ci\+i^\{\+\}=\\arg\\max\_\{1\\leq i\\leq t\}c\_\{i\}^\{\+\}andi−=arg⁡max1≤i≤t⁡ci−i^\{\-\}=\\arg\\max\_\{1\\leq i\\leq t\}c\_\{i\}^\{\-\}denote the iteration indices yielding the highest supporting and refuting confidence scores, respectively\. Max\-pooling captures the strongest supporting and refuting signals across all iterations, whileαt\+⁣∗\\alpha\_\{t\}^\{\+\*\}andαt−⁣∗\\alpha\_\{t\}^\{\-\*\}track the corresponding best arguments\. The best arguments are passed to the Controller for adjudication, ensuring that the final verdict is based on the strongest arguments rather than those from the most recent iteration alone\.

### 3\.4Meta\-Controller

The Controller implements a meta\-reflection mechanism that determines whether the current evidence is sufficient for a final decision or whether further reasoning is required\.

#### Confidence Gap\.

We use the confidence gap between the dialectical perspectives as the control signal:

Δt=\|Ct\+−Ct−\|\.\\Delta\_\{t\}=\|C\_\{t\}^\{\+\}\-C\_\{t\}^\{\-\}\|\.\(8\)A larger gap indicates more consistent evidence toward one verdict, while a smaller gap suggests ambiguity\.

#### Meta\-Reflective Decision\.

The Controller employs an LLM\-based adjudication function that determines whether to stop reasoning or continue exploration:

Φ⁡\(αt\+⁣∗,αt−⁣∗,Δt,τ\)=\{Adj⁡\(⋅\),Δt≥τ,Continue,otherwise\.\\Phi\(\\alpha\_\{t\}^\{\+\*\},\\alpha\_\{t\}^\{\-\*\},\\Delta\_\{t\},\\tau\)=\\begin\{cases\}\\mathrm\{Adj\}\(\\cdot\),&\\Delta\_\{t\}\\geq\\tau,\\\\ \\textsc\{Continue\},&\\text\{otherwise\}\.\\end\{cases\}\(9\)
Here,Adj⁡\(⋅\)\\mathrm\{Adj\}\(\\cdot\)denotes an LLM\-based adjudicator that takes as input both the best supporting argumentαt\+⁣∗\\alpha\_\{t\}^\{\+\*\}and the best refuting argumentαt−⁣∗\\alpha\_\{t\}^\{\-\*\}, and produces a final decision by comparing their consistency with the accumulated evidence\. The adjudicator synthesizes the two dialectical perspectives in a structured prompt and outputs the predicted labely^∈\{Sup,Ref\}\\hat\{y\}\\in\\\{\\textsc\{Sup\},\\textsc\{Ref\}\\\}\.

If the maximum iterationTTis reached, the controller directly outputs the final verdict using the best available arguments:

y^=Adj⁡\(αT\+⁣∗,αT−⁣∗,θC\)\.\\hat\{y\}=\\mathrm\{Adj\}\(\\alpha\_\{T\}^\{\+\*\},\\alpha\_\{T\}^\{\-\*\};\\theta\_\{C\}\)\.
This design enables adaptive termination based on the discriminative strength of dialectical evidence, balancing efficiency and reasoning completeness\.

Algorithm 1Dialectical Agentic REasoning0:Claim

ss, KG

𝒢=\(ℰ,ℛ,𝒯\)\\mathcal\{G\}=\(\\mathcal\{E\},\\mathcal\{R\},\\mathcal\{T\}\), threshold

τ\\tau, maximum iterations

TT
1:Initialize:

C0\+←0C\_\{0\}^\{\+\}\\leftarrow 0,

C0−←0C\_\{0\}^\{\-\}\\leftarrow 0,

E1:0←∅E\_\{1:0\}\\leftarrow\\emptyset,

α\+⁣∗←∅\\alpha^\{\+\*\}\\leftarrow\\emptyset,

α−⁣∗←∅\\alpha^\{\-\*\}\\leftarrow\\emptyset
2:Extract entity set

ℰs\\mathcal\{E\}\_\{s\}from

ssvia entity linking

3:for

t=1,2,…,Tt=1,2,\\ldots,Tdo

4:Retriever:Retrieve adjacency set

Γs\(t\)←⋃e∈ℰs\{\(h,r,o\)∈𝒯∣h=e∨o=e\}\\Gamma\_\{s\}^\{\(t\)\}\\leftarrow\\bigcup\_\{e\\in\\mathcal\{E\}\_\{s\}\}\\\{\(h,r,o\)\\in\\mathcal\{T\}\\mid h=e\\lor o=e\\\}
5:Retriever:Filter relevant evidence

E\(t\)←GE​\(Γs\(t\),s,θE\)E^\{\(t\)\}\\leftarrow G\_\{E\}\(\\Gamma\_\{s\}^\{\(t\)\},s;\\theta\_\{E\}\)
6:Update accumulated evidence:

E1:t←E1:t−1∪E\(t\)E\_\{1:t\}\\leftarrow E\_\{1:t\-1\}\\cup E^\{\(t\)\}
7:Reasoner:Generate supporting argument and score

ct\+,αt\+←D\+\(E1:t,s;θD\+\)c\_\{t\}^\{\+\},\\alpha\_\{t\}^\{\+\}\\leftarrow D^\{\+\}\(E\_\{1:t\},s;\\theta\_\{D\}^\{\+\}\)
8:Reasoner:Generate refuting argument and score

ct−,αt−←D−\(E1:t,s;θD−\)c\_\{t\}^\{\-\},\\alpha\_\{t\}^\{\-\}\\leftarrow D^\{\-\}\(E\_\{1:t\},s;\\theta\_\{D\}^\{\-\}\)
9:Update cumulative confidences:

Ct\+←max⁡\(Ct−1\+,ct\+\),Ct−←max⁡\(Ct−1−,ct−\)C\_\{t\}^\{\+\}\\leftarrow\\max\(C\_\{t\-1\}^\{\+\},c\_\{t\}^\{\+\}\),\\quad C\_\{t\}^\{\-\}\\leftarrow\\max\(C\_\{t\-1\}^\{\-\},c\_\{t\}^\{\-\}\)
10:Update best arguments:

i\+←arg⁡max1≤i≤t​ci\+,α\+⁣∗←αi\+\+;i−←arg⁡max1≤i≤t​ci−,α−⁣∗←αi−−i^\{\+\}\\leftarrow\\arg\\max\_\{1\\leq i\\leq t\}c\_\{i\}^\{\+\},\\quad\\alpha^\{\+\*\}\\leftarrow\\alpha\_\{i^\{\+\}\}^\{\+\};\\quad i^\{\-\}\\leftarrow\\arg\\max\_\{1\\leq i\\leq t\}c\_\{i\}^\{\-\},\\quad\\alpha^\{\-\*\}\\leftarrow\\alpha\_\{i^\{\-\}\}^\{\-\}
11:Compute confidence gap:

Δt←\|Ct\+−Ct−\|\\Delta\_\{t\}\\leftarrow\|C\_\{t\}^\{\+\}\-C\_\{t\}^\{\-\}\|
12:if

Δt≥τ\\Delta\_\{t\}\\geq\\tauthen

13:Controller:

y^←Adj⁡\(α\+⁣∗,α−⁣∗,θC\)\\hat\{y\}\\leftarrow\\mathrm\{Adj\}\(\\alpha^\{\+\*\},\\alpha^\{\-\*\};\\theta\_\{C\}\)
14:return

y^\\hat\{y\}
15:endif

16:Expand entity set:

ℰs←ℰs∪Entities⁡\(E\(t\)\)\\mathcal\{E\}\_\{s\}\\leftarrow\\mathcal\{E\}\_\{s\}\\cup\\mathrm\{Entities\}\(E^\{\(t\)\}\)
17:endfor

18:Controller:

y^←Adj⁡\(α\+⁣∗,α−⁣∗,θC\)\\hat\{y\}\\leftarrow\\mathrm\{Adj\}\(\\alpha^\{\+\*\},\\alpha^\{\-\*\};\\theta\_\{C\}\)
19:return

y^\\hat\{y\}

### 3\.5Overall Algorithm

Algorithm[1](https://arxiv.org/html/2609.13808#alg1)summarizes the complete DARE framework\. The procedure follows an iterative retrieve–reason–reflect paradigm, where evidence is progressively expanded and refined across iterations\. Early iterations typically resolve simple cases \(e\.g\., single\-hop evidence suffices\), whereas more complex claims require multiple rounds of retrieval and dialectical reasoning to accumulate discriminative evidence\.

A key component of the framework is the dynamic expansion of the entity set based on newly retrieved triples, enabling multi\-hop exploration over the knowledge graph\. This mechanism allows the retriever to progressively traverse beyond initially mentioned entities, thereby supporting compositional reasoning over longer inference chains\.

The iteration process is bounded by a maximum stepTTto ensure computational efficiency\. In practice, the controller typically converges within a small number of iterations \(e\.g\., 2–3 steps\), as the confidence gap criterion often triggers early termination when evidence becomes sufficiently discriminative\.

## 4Experiments

### 4\.1Experimental Setup

#### Dataset\.

We evaluate DARE on the FactKG dataset\(kim2023factkg\), a benchmark designed for fact verification on knowledge graphs\. FactKG contains 108,764 claims based on factual information from the DBpedia knowledge graph\(lehmann2015dbpedia\)\. Each instance consists of a claim, the corresponding DBpedia entities, a binary label \(Supported or Refuted\), and a reasoning type\. The test set comprises 9,041 claims spanning five reasoning types: one\-hop \(1,914\), conjunction \(3,069\), existence \(870\), multi\-hop \(1,874\), and negation \(1,314\), providing a comprehensive evaluation across varying claim complexities\.

#### Baselines\.

Following the evaluation protocol ofkim2023factkgandhao2025pgr, we categorize baselines into two groups based on evidence usage:

Without Evidence\.These methods rely solely on the claim text: BERT\(devlin2019bert\)and BlueBERT\(bluebert2019\)are fine\-tuned on the FactKG training set, whereas Flan\-T5\(chung2024flant5\)is evaluated in the zero\-shot setting\. ChatGPT uses 12\-shot prompting\.

With Evidence\.These methods incorporate KG evidence: GEAR\(zhou2019gear\)is a full\-training model optimized for FactKG\. KG\-GPT\(kim2023kggpt\)performs LLM\-based KG reasoning under few\-shot settings; we report results for the original gpt\-3\.5\-turbo variant \(KG\-GPT\), GPT\-4o variant \(KG\-GPT†\), GPT\-4o\-mini variant \(KG\-GPT∗\), and DeepSeek\-V3 variant \(KG\-GPTd\)\. ProgramFC\(pan2023programfc\)generates verification programs\. PGR\(hao2025pgr\)is the strongest program\-generation baseline; we report PGR \(GPT\-4o\), PGR\-mini \(GPT\-4o\-mini\), and PGRd\(DeepSeek\-V3\)\.

#### Implementation Details\.

We implement DARE using Qwen3\-8B\(qwen2025qwen3\)as the backbone LLM for all three agents\. All agents share the same model but are instantiated with distinct role\-specific prompts\. The overall framework operates in a zero\-shot setting without task\-specific demonstrations\.

The Retriever performs semantic filtering over retrieved KG triples using zero\-shot prompting\. The Reasoner is instructed with dialectical prompt that guides it to evaluate the evidence from both supporting and refuting perspectives and produce corresponding confidence scores in a unified output\. The Controller uses an adjudication prompt that takes both dialectical arguments as input and outputs a final verdict\. The confidence gap threshold is set toτ=0\.5\\tau=0\.5, and the maximum number of iterations isT=5T=5\. The decoding hyperparameters are temperature = 0\.1 and top\-p = 0\.9\.

### 4\.2Main Results

Table 1:Accuracy \(%\) on the FactKG dataset\. Methods are categorized by evidence usage and training strategy\. The best results are inbold\. DARE uses Qwen3\-8B as the backbone model\.Table[1](https://arxiv.org/html/2609.13808#S4.T1)presents the experimental results\. DARE achieves the highest overall accuracy \(88\.12%\), outperforming PGR with GPT\-4o \(86\.82%\) by 1\.30% despite using a model with substantially fewer parameters\. This gap is primarily driven by two claim types where dialectical and iterative reasoning offer the most leverage\.

On negation claims, DARE reaches 88\.89%, a 7\.73% improvement over PGR’s 81\.16%\. Program\-generation methods like PGR tend to map natural language negations directly into programmaticNOToperators, searching for missing edges rather than resolving the underlying logic\. In contrast, the dialectical reasoner explicitly evaluates both supporting and refuting perspectives, forcing the model to confront potential contradictions instead of defaulting to a single\-direction interpretation\. A similar pattern emerges on existence claims \(95\.23% vs\. 93\.10%\), where the retrieve\-then\-filter strategy ensures that all considered relations are grounded in the KG, eliminating the relation mapping errors that degrade program\-generation baselines\.

On multi\-hop claims, the most challenging type, DARE achieves 77\.28%, surpassing PGR’s 75\.96%\. Multi\-hop verification requires chaining multiple relational paths, and the iterative dialectical process enables the reasoner to accumulate and cross\-validate intermediate evidence at each step before converging to a judgment\. Across five reasoning types, DARE outperforms PGR on four, achieving these gains under a 0\-shot setting without any demonstrations—suggesting that structured agentic reasoning can effectively compensate for the absence of in\-context examples\.

### 4\.3Ablation Study

To understand the contribution of each agent, we conduct ablation experiments by removing the core mechanism of each agent respectively\.

Table 2:Ablation study on FactKG\. Each variant removes the core mechanism of one agent\.Δ\\Deltashows the accuracy change relative to the full model\.#### Reasoner’s Dialectical Reasoning\.

Removing dialectical reasoning \(i\.e\., using only a single\-perspective evaluation instead of bidirectional scoring\) causes the largest accuracy drop \(−\-24\.92%\)\. This dramatic decline confirms that the dialectical structure is essential for calibrating confidence and combating self\-evaluation distortion\. Without the opposing perspective, the model systematically overestimates supporting evidence, leading to a strong bias toward the*Supported*verdict and severely undermining verification reliability\.

#### Controller’s Meta\-Reflection\.

Removing meta\-reflection \(i\.e\., fixing the framework to a single iteration without dynamic termination\) reduces accuracy by 18\.01%\. This confirms that iterative evidence gathering with confidence\-driven termination is critical: complex claims such as multi\-hop and negation require multiple rounds of evidence expansion and dialectical re\-evaluation, which a single\-pass pipeline cannot provide\.

#### Retriever’s Semantic Filtering\.

Removing semantic filtering \(i\.e\., passing all adjacent triples directly to the Reasoner\) reduces accuracy by 15\.91%\. While full adjacency retrieval prevents relation mapping failures, the resulting noise from irrelevant relations severely degrades the Reasoner’s evaluation quality\. This highlights that precise evidence selection is as important as comprehensive retrieval—the Retriever must filter out irrelevant relations to enable effective dialectical reasoning\.

### 4\.4Analysis

#### Effect of Confidence Gap Threshold\.

Figure[4](https://arxiv.org/html/2609.13808#S4.F4)shows the effect of varying the confidence gap thresholdτ\\tau\. Whenτ\\tauis too low \(e\.g\., 0\.3\), the framework terminates prematurely, missing the opportunity to gather additional evidence for complex claims\. Whenτ\\tauis too high \(e\.g\., 0\.6\), the framework requires excessively strong confidence before terminating, leading to unnecessary iterations that may introduce noise\. The optimal thresholdτ=0\.5\\tau=0\.5balances early termination for easy claims with sufficient exploration for complex ones\.

#### Effect of Maximum Iterations\.

Figure[3](https://arxiv.org/html/2609.13808#S4.F3)shows the effect of varying the maximum iteration limitTT\. Accuracy improves substantially fromT=3T=3\(80\.60%\) toT=5T=5\(88\.12%\), as additional iterations enable the framework to accumulate evidence for complex claims\. However,T=6T=6yields slightly lower accuracy \(87\.96%\), suggesting that excessive iterations introduce noise through unnecessary entity expansion\. We setT=5T=5as the default, which achieves the best performance while maintaining reasonable efficiency\.

Figure 3:Impact of the maximum iteration limits \(TT\) on the overall accuracy\.Figure 4:Impact of the confidence gap threshold \(τ\\tau\) on the overall accuracy\.
#### Case Study\.

Table[3](https://arxiv.org/html/2609.13808#S4.T3)illustrates how DARE handles multi\-hop claims that require evidence from multiple KG triples\. The claim “Aaron Boogaard was from Canada where one of the languages spoken is French” involves a conjunction of two sub\-claims: a birth\-location fact and a language fact\.

PGR \(GPT\-4o\) generates a program that attempts to match both relations in a single execution, but fails because the strict programmatic syntax cannot map “from” to the exact valid KG paths, resulting in an execution failure and a*False*prediction\. In contrast, DARE resolves this claim over two iterations\. In the first iteration, the retriever extracts\(Aaron\_Boogaard, placeOfBirth, Canada\), which supports the birth\-location sub\-claim but leaves the language sub\-claim unverified\. The reasoner assignsc\+=0\.55c^\{\+\}=0\.55andc−=0\.45c^\{\-\}=0\.45, yieldingΔ=0\.10<τ\\Delta=0\.10<\\tau—insufficient to terminate\. The controller then expands the entity set to include Canada, enabling the retriever in the second iteration to find\(Canada, language, French\_language\)\. With both sub\-claims now grounded, the reasoner updates its confidence toc\+=0\.95c^\{\+\}=0\.95andc−=0\.10c^\{\-\}=0\.10, producingΔ=0\.85≥τ\\Delta=0\.85\\geq\\tau\. The controller adjudicates and outputs*True*\. This case demonstrates how iterative evidence expansion combined with confidence\-driven termination allows DARE to accumulate evidence across hops, avoiding the single\-pass execution failures of program\-generation baselines\.

Claim:Aaron Boogaard was from Canada where one of the languages spoken is French\.Label:TrueType:Conjunction / Multi\-claimBaseline: PGR \(GPT\-4o\)MATCH\(Aaron\_Boogaard, placeOfBirth, ?x\) AND MATCH\(?x, language, French\_language\)\.Error:The strict programmatic syntax fails to map "from" to the exact valid KG paths in a single pass, leading to an execution failure\.Prediction:False×\\timesDARE \(Ours\)Iter\. 1:Retriever:Extracts\(Aaron\_Boogaard, placeOfBirth, Canada\)\.Reasoner:c\+=0\.55c^\{\+\}=0\.55\(Supports birth location, but missing language evidence\),c−=0\.45c^\{\-\}=0\.45\(Refutes due to unverified language claim\)\.Controller:Δ=0\.10<τ\\Delta=0\.10<\\tau\. Continues reasoning\.Iter\. 2:Retriever:Expands via ‘Canada‘ and extracts\(Canada, language, French\_language\)\.Reasoner:Cumulative evidence now satisfies both sub\-claims\.c\+=0\.95c^\{\+\}=0\.95,c−=0\.10c^\{\-\}=0\.10\.Controller:Δ=0\.85≥τ\\Delta=0\.85\\geq\\tau\. Terminates and adjudicates\.Prediction:True✓\\checkmarkTable 3:Case study on a multi\-hop claim\.

## 5Conclusion

We have presented DARE, an agentic system for structured knowledge fact checking centered on the observation that LLMs suffer from self\-evaluation distortion—a systematic bias toward overestimating supporting evidence when assessing claims from a single perspective\. DARE mitigates this through dialectical bidirectional reasoning, which forces the model to argue from both supporting and refuting positions, producing calibrated confidence estimates rather than one\-sided judgments\. A confidence\-gap\-driven controller then determines whether evidence is sufficiently discriminative or whether the agents should continue exploring, enabling dynamic self\-correction absent in single\-pass methods\. On FactKG, DARE with Qwen3\-8B achieves state\-of\-the\-art accuracy \(88\.12%\), surpassing PGR with GPT\-4o while using a model substantially smaller\. These findings indicate that addressing self\-evaluation distortion through structured dialectical evaluation is a promising direction for improving LLM\-based verification, and more broadly, that agentic designs which confront models with opposing evidence can yield more reliable reasoning over structured knowledge\. We believe this principle extends beyond fact verification to other tasks requiring calibrated judgment over structured data, such as multi\-hop reasoning and question answering\.

## Limitations

Despite the demonstrated effectiveness of DARE in structured knowledge fact checking, two limitations warrant further discussion\. First, the Retriever’s adjacency retrieval extracts all neighboring triples for each entity, which scales with the degree of the entity\. For extremely dense knowledge graphs where hub entities possess thousands of adjacent relations, the initial retrieval step could introduce overhead, though semantic filtering subsequently reduces the set passed to downstream agents\. Second, our empirical evaluation focuses on fact checking grounded in structured evidence\. The results demonstrate the effectiveness of DARE in this setting, while its extension to a broader range of reasoning tasks holds substantial potential\. Realizing this potential across diverse reasoning tasks represents an important direction for future work\.

## References

## Appendix ADataset Details

In this work, we conduct all experiments on the FactKG dataset, a challenging knowledge graph reasoning benchmark that covers diverse reasoning types and linguistic styles\. The dataset contains two main linguistic categories:writtenformal sentences andcolloquialinformal sentences, where the colloquial set is further divided into model\-generated colloquial texts and presupposition\-based colloquial texts\.

To comprehensively illustrate the data composition, we summarize the detailed data statistics across five typical reasoning types, including one\-hop reasoning, conjunction reasoning, existence reasoning, multi\-hop reasoning, and negation reasoning\. Table[4](https://arxiv.org/html/2609.13808#A2.T4)presents the overall data distribution and detailed statistics\. Figure[5](https://arxiv.org/html/2609.13808#A2.F5)shows the graph patterns adopted by conjunction and multi\-hop claims\.

## Appendix BDetails of Baseline Methods

#### Baselines\.

Following the evaluation protocol ofkim2023factkgandhao2025pgr, we categorize baselines into two groups based on evidence usage\. We elaborate the core design and implementation of each baseline as follows\.

- •BERT\(devlin2019bert\): A classic bidirectional Transformer model pre\-trained on large\-scale text corpora\. We adopt the standard architecture and train it as a text classifier via full training, which takes only claim text as input without external knowledge evidence\.
- •BlueBERT\(bluebert2019\): A domain\-enhanced variant derived from BERT, further pre\-trained on biomedical texts to strengthen semantic representation capability\. Consistent with the unified setting, we train it with the full\-training strategy and use pure claim text for prediction\.
- •Flan\-T5\(chung2024flant5\): A large language model optimized via multi\-task instruction tuning\. Following the original FactKG evaluation protocol, we evaluate Flan\-T5 in the zero\-shot setting using only the claim text\.
- •GEAR\(zhou2019gear\): A graph neural network based framework specially optimized for the FactKG benchmark\. It leverages knowledge graph evidence and is trained under the full\-training paradigm to conduct fact verification\.
- •KG\-GPT\(kim2023kggpt\): A typical LLM\-based knowledge graph reasoning method working under few\-shot settings\. It combines large language models with retrieved KG evidence to complete reasoning\. We evaluate multiple model variants, including the GPT\-3\.5\-turbo version \(KG\-GPT\), GPT\-4o version \(KG\-GPT†\), GPT\-4o\-mini version \(KG\-GPT∗\) and DeepSeek\-V3 version \(KG\-GPTd\)\.
- •ProgramFC\(pan2023programfc\): An interpretable fact verification approach that converts natural language claims into executable logical programs\. It verifies factual correctness by executing the generated programs over knowledge graph data\.
- •PGR\(hao2025pgr\): A state\-of\-the\-art program generation baseline for KG\-based fact verification\. We test three variants built upon different backbones, namely PGR \(GPT\-4o\), PGR\-mini \(GPT\-4o\-mini\) and PGRd\(DeepSeek\-V3\)\.

![Refer to caption](https://arxiv.org/html/2609.13808v1/Dataset.png)Figure 5:Graph patterns for conjunction and multi\-hop claims in FactKG\.Table 4:FactKG Dataset Statistics

## Appendix CCross\-Dataset Generalization

We further evaluate DARE on Fact\-FOLX\-KG\(hao\-wu\-2026\-extending\)to assess its transferability across datasets\. Fact\-FOLX\-KG is a structured knowledge fact checking benchmark that requires first\-order logic \(FOL\) reasoning across five categories\.

Table 5:Accuracy \(%\) on the Fact\-FOLX\-KG dataset\. A superscript∗\*denotes a DeepSeek\-V3\.2\-based baseline variant; unmarked LLM\-based baseline variants use GPT\-4o\. For our method, we report results obtained with Qwen3\-8B\. The best results among the included methods are inbold\.
## Appendix DPrompt Templates

We provide the detailed prompt templates used in DARE\.

Prompt: Dialectical Reasoning\#\# Instruction: You are aFact\-Checking Expert\. Your task is to verify a ’Claim’ using provided ’Retrieved Facts’ from a Knowledge Graph and your own internal knowledge and provide two independent confidence scores \(0\.0 to 1\.0\) for a given claim\. \#\# GUIDELINES: 1\.Relevance Check: The ’Retrieved Facts’ are candidates and may be IRRELEVANT to the claim\. If they are irrelevant, rely on your internal knowledge\. 2\.Conflict Detection: If a Retrieved Fact explicitly contradicts the claim \(e\.g\., Claim says ’born in A’, Fact says ’born in B’\), the verdict is FALSE\. 3\.Supportive Evidence: If a Retrieved Fact confirms the claim, the verdict is TRUE\. 4\.Negation: Pay attention to "not", "no", etc\. If the claim is negated, and the facts support the positive statement, the verdict is FALSE\. \#\# EVALUATION LOGIC: 1\.True\_Confidence: How strongly does the Evidence and your knowledge support the claim being TRUE? 2\.False\_Confidence: How strongly does the Evidence \(especially entity mismatches or relation conflicts\) suggest the claim is FALSE? \#\# CRITICAL ADVERSARIAL CHECK: \- If the claim says ’Object A’ but evidence shows ’Object B’ for the fact, False\_Confidence should be very high\. \#\# FORMAT: Reasoning: <Step\-by\-step analysis\> Verdict: \[TRUE or FALSE\] \- Scores: \{"True\_Confidence": float, "False\_Confidence": float\}