PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Comprehensive Stage-wise Benchmarking of Large Language Models in Fact-Checking

Hongzhan Lin, Zixin Chen, Zhiqi Shen, Ziyang Luo, Zhen Ye, Jing Ma, Tat-Seng Chua, Guandong Xu

arXiv:2601.02669v1cs.CL

TL;DR

Evaluations often center on claim verification, while assessing open-form outputs across all fact-checking stages remains challenging. FactArena benchmarks LLMs through a stage-wise arena and adaptive claim evolution; across 16 models, it produces robust rankings and reveals discrepancies static accuracy metrics fail to capture.

  • Problem

    Existing evaluations often center on claim verification, while assessing open-form outputs across all fact-checking stages and integrating them into an overall competence measure remains challenging.

    Source: C. Arena-styled Stage-wise Judgment · p. 4, §V · p. 10

  • Method

    FactArena combines stage-wise arena evaluation of claim decomposition, evidence retrieval, and verdicts with an adaptive claim-evolution mechanism.

    Source: §I · p. 2

  • Results

    Across 16 models, FactArena produces robust and fair rankings and reveals performance discrepancies that static accuracy metrics fail to capture.

    Source: §V · p. 10

  • Takeaways & Limitations

    Static accuracy metrics alone fail to capture performance discrepancies, supporting assessment of LLM fact-checking competence across the full pipeline.

    Source: §V · p. 10

Abstract

from arXiv · show

Large Language Models (LLMs) are increasingly deployed in real-world fact-checking systems, yet existing evaluations focus predominantly on claim verification and overlook the broader fact-checking workflow, including claim extraction and evidence retrieval. This narrow focus prevents current benchmarks from revealing systematic reasoning failures, factual blind spots, and robustness limitations of modern LLMs. To bridge this gap, we present FactArena, a fully automated arena-style evaluation framework that conducts comprehensive, stage-wise benchmarking of LLMs across the complete fact-checking pipeline. FactArena integrates three key components: (i) an LLM-driven fact-checking process that standardizes claim decomposition, evidence retrieval via tool-augmented interactions, and justification-based verdict prediction; (ii) an arena-styled judgment mechanism guided by consolidated reference guidelines to ensure unbiased and consistent pairwise comparisons across heterogeneous judge agents; and (iii) an arena-driven claim-evolution module that adaptively generates more challenging and semantically controlled claims to probe LLMs' factual robustness beyond fixed seed data. Across 16 state-of-the-art LLMs spanning seven model families, FactArena produces stable and interpretable rankings. Our analyses further reveal significant discrepancies between static claim-verification accuracy and end-to-end fact-checking competence, highlighting the necessity of holistic evaluation. The proposed framework offers a scalable and trustworthy paradigm for diagnosing LLMs' factual reasoning, guiding future model development, and advancing the reliable deployment of LLMs in safety-critical fact-checking applications.

I. INTRODUCTION

FactArena is an automated arena-style framework for evaluating LLMs across the full fact-checking pipeline, including claim extraction, evidence retrieval, and claim verification. It combines stage-wise evaluation, judgments aligned across diverse evaluators, and adaptive claim generation to assess factual reasoning and robustness.

I. INTRODUCTION

  • Traditional LLM fact-checking evaluations primarily assess claim verification, overlooking earlier stages that can compromise the integrity of end-to-end systems.The paper argues that systematic assessment of every stage is essential as LLMs are used on complex real-world claims, while open-ended judgments and annotation costs make evaluation challenging.Source: §I · p. 1
  • FactArena combines holistic pipeline coverage, consensus across heterogeneous evaluators, and automation to enable fairer, more reliable, large-scale assessment.The framework draws on collective intelligence to reduce reliance on any single judgment and modularizes auditing to limit costly manual annotation.Source: §I · p. 2
  • FactArena evaluates LLMs in stage-wise peer battles for claim decomposition and planning, tool-augmented evidence retrieval and integration, and explanation quality and verdict soundness.The framework presents this as a scalable, automated approach to understanding models’ strengths and limitations across the pipeline.Source: §I · p. 2
  • FactArena’s arena-driven claim evolution adaptively generates increasingly challenging, unseen claims to probe LLM factual robustness and knowledge boundaries beyond fixed test sets.Claims are generated based on model performance, supporting adaptive, model-centric evaluation.Source: §I · p. 2
  • The introduction reports that diverse evaluator agents’ judgments are strongly consistent and align closely with human expert assessments, providing actionable insights for LLM auditing.The authors also state that experiments show FactArena improves the reliability and wholeness of fact-checking workflow evaluation.Source: §I · p. 2

II. RELATED WORK … B. LLM-based Fact-Checking Task Formulation

The paper argues that static, verification-focused evaluations miss the open-ended demands of fact-checking and proposes a comprehensive, arena-style evaluation of LLMs across the full pipeline. FactArena operationalizes that evaluation through claim extraction, evidence retrieval, and justification-based verdict prediction.

II. RELATED WORK

  • Prior static benchmarks rely on expert-designed datasets and specialized tasks, limiting adaptability to emerging misinformation and the open-ended complexity of real-world fact-checking.The paper instead uses LLM-generated justifications for dynamic auditing of limitations beyond veracity prediction.Source: §II · p. 2
  • Prior multi-agent audits of fact-checking largely focus on claim verification, leaving claim extraction and evidence retrieval insufficiently evaluated, especially where no closed-form ground truth exists.Source: §II · p. 2

A. Overview

  • FactArena aims to provide comprehensive, unbiased arena-style evaluation of LLMs’ stage-wise abilities to fact-check complex claims.Its components construct an LLM-driven fact-checking process, compare models through stage-wise peer battles and automatic judgment, and generate progressively challenging claims based on model performance.Source: A. Overview · p. 3

B. LLM-based Fact-Checking Task Formulation

  • The proposed pipeline uses the target LLM as its core controller and runs three sequential stages: claim extraction, evidence retrieval, and justification production with verdict prediction.Source: B. LLM-based Fact-Checking Task… · p. 3
  • Claim extraction decomposes a complex claim into verifiable sub-claims that guide subsequent evidence retrieval and justification generation.The number of sub-claims is denoted by k.Source: B. LLM-based Fact-Checking Task… · p. 3
  • For evidence retrieval, the target LLM uses Google Search on the original claim and receives the top-ranked result’s title and snippet as context.It processes that information with the claim and decomposed sub-claims to extract and summarize salient facts as evidence supporting or refuting the claim.Source: B. LLM-based Fact-Checking Task… · p. 3
  • In the final stage, the LLM uses the original claim, sub-claims, and retrieved evidence to produce a justification and predicted verdict.The justification is denoted r and the verdict y.Source: B. LLM-based Fact-Checking Task… · p. 3

C. Arena-styled Stage-wise Judgment

FactArena compares anonymous models through stage-wise peer battles and uses shared judging guidelines to evaluate open-form outputs across the fact-checking pipeline. Pairwise judgments are converted into model rankings using Elo and Bradley–Terry methods.

Arena-styled Stage-wise Judgment

  • Randomly selected anonymous model pairs generate stage-wise outputs for peer battles, enabling comparison of open-form responses and rankings of relative fact-checking capability.The evaluated outputs include sub-claims, evidence, and verification responses.Source: C. Arena-styled Stage-wise Judgment · p. 4
  • Guidelines are established for all three open-form stages, with judges drawn from different model families and barred from evaluating their own responses.The judge panel is diversified to mitigate bias.Source: C. Arena-styled Stage-wise Judgment · p. 4
  • For claim extraction, judges iteratively synthesize competing sub-claims into a shared guideline, rotating judges and samples until all candidate decompositions are incorporated.The consolidation aims to capture different decomposition strengths and provide a consistent basis for evaluation.Source: C. Arena-styled Stage-wise Judgment · p. 4
  • For evidence retrieval, judges use Wikipedia searches and gold evidence as a factual reference, not a golden label, to assess whether outputs capture essential claim-supporting or refuting content.The reference is intended to mitigate factual errors in judge assessments.Source: C. Arena-styled Stage-wise Judgment · p. 4
  • Verification verdicts are evaluated as classifications, while justifications are judged for helpfulness, informativeness, soundness, and readability, [46],.Justification assessment checks verdict consistency, inclusion of salient evidence, logical coherence, and clear articulation.Source: C. Arena-styled Stage-wise Judgment · pp. 4–5
  • Elo ratings aggregate pairwise wins but depend on matchup order; Bradley–Terry estimates order-independent strengths, splits ties into half-wins, and sorts ratings into final rankings, [49].Bradley–Terry is used alongside Elo to mitigate ranking sensitivity to sequential matchup updates.Source: C. Arena-styled Stage-wise Judgment · p. 5

D. Arena-driven Claim Evolution

FactArena uses arena feedback to evolve claims when models succeed, probing their factual reasoning beyond fixed test sets and helping identify scenarios where they perform poorly.

D. Arena-driven Claim Evolution

  • Fixed fact-checking test sets offer only superficial assessments, obscure deeper limitations and knowledge boundaries, and may risk test leakage.The section motivates harder claims as a way to probe model underperformance more comprehensively.Source: D. Arena-driven Claim Evolution · p. 5
  • When all target models correctly judge a claim, FactArena reverses its semantics, then iteratively evolves equivalent, harder variants if they also correctly judge the contrastive claim.The iterative evolution is driven by arena battle judgments.Source: D. Arena-driven Claim Evolution · p. 5
  • Each evolved claim is evaluated by judge agents against target LLMs, with results stored in a judgment pool to identify scenarios where models perform poorly.This process yields insights into model capacities across diverse, complex fact-checking claims.Source: D. Arena-driven Claim Evolution · p. 5

IV. EXPERIMENTS · A. Experimental Settings · B. Implementation Details

FactArena evaluates complex claims from HOVER and FEVEROUS across 16 LLMs, using controlled sampling, arena judgments, and claim evolution to structure its experiments.

A. Experimental Settings

  • Experiments cover complex HOVER and FEVEROUS claims requiring multi-hop reasoning beyond simple factual matching.The paper refers readers to Table I for detailed dataset statistics; those values are not provided here.Source: A. Experimental Settings · p. 5
  • The evaluation spans 16 LLMs across 7 model families, with DeepSeek-V3, GPT-o4 mini, Gemini 2.5 Flash, and Qwen3 (235B) selected as judge agents.The judge panel was selected by balancing performance and computational cost.Source: A. Experimental Settings · p. 6

B. Implementation Details

  • GPT-o4 mini controls arena-driven claim evolution by reversing semantics to create harder claims, and target models and judges use temperature 0.0.The temperature setting is intended to make evaluation results as reproducible as possible.Source: B. Implementation Details · p. 6
  • 400 claims—200 from each dataset—were sampled, with 8 target-model responses per claim per task for pairwise comparisons.The sampling was intended to diversify comparisons while keeping the total number of battles controllable, following combinatorial coverage theory.Source: B. Implementation Details · p. 6
  • 85 of the 400 claims were correctly predicted by all selected target models and were semantically reversed and evolved.The final arena-styled judgment produced about 13,000 valid judgments; each target LLM participated in approximately 1,600 comparisons, averaging roughly 104 comparisons per model pair.Source: B. Implementation Details · p. 6
  • Compared results were averaged over three random runs, and evaluating one target model cost about $15 and 3 hours.Source: B. Implementation Details · p. 6

C. Main Results · D. Analysis of Judgment Reliability with Guideline

FactArena’s stage-wise results show that overall fact-checking rankings and accuracy-only scores can tell different stories. Its guideline-based judging produces more consistent, less biased assessments that align best with human annotations.

C. Main Results

  • GPT-o3 and DeepSeek-R1 rank first and second by Elo, respectively, consistently outperforming other models across all fact-checking stages.The overall ranking is determined jointly by evaluator-panel votes.Source: C. Main Results · p. 6
  • Models generally perform consistently across stages, but Gemini 2.5 Pro surpasses GPT-4.5 in claim extraction and verification while trailing it in evidence retrieval.Llama 4 Maverick and Gemini 2.5 Flash show slight instability specifically in claim extraction.Source: C. Main Results · p. 6
  • Gemini 2.5 Pro’s highest claim-verification accuracy advantage diminishes across the full pipeline, while Llama 4 Maverick and GPT-4.1 show similar inconsistencies.The passage concludes that claim-verification accuracy alone is insufficient to represent overall fact-checking capability.Source: C. Main Results · p. 7

D. Analysis of Judgment Reliability with Guideline

  • FactArena produces more robust, stratified Elo rankings; guidelines diminish Gemini 2.5 Flash and DeepSeek-V3 preferences for same-family models.Gemini 2.5 Flash also deviates substantially from other judges, especially when evaluating lower-ranked models.Source: D. Analysis of Judgment Reliability… · p. 7, Figure 4 · p. 7
  • Guidelines improve average judge accuracy in claim extraction and justification production, while evidence-retrieval accuracy remains relatively unchanged.The guidelines generally reduce subjective evaluator bias across the pipeline.Source: D. Analysis of Judgment Reliability… · p. 7, Table III · p. 7
  • Inter-judge consistency across the three ablation settings further supports the reliability of guideline-driven evaluation.The reported alignment indicates more consistent, human-aligned evaluations of the fact-checking workflow.Source: D. Analysis of Judgment Reliability… · p. 8, Table V · p. 8

E. Analysis of Arena-driven Claim Evolution · F. Qualitative Comparison with Traditional Benchmark

Arena-driven claim evolution made claims harder over successive rounds, while human experts preferred FactArena judges’ explanations for their information and quality. The analyses also report model-rating changes under evolution and compare the evaluation designs using expert judgments.

E. Analysis of Arena-driven Claim Evolution

  • The ablation compared FactArena with claim evolution against variants that reverse trivial claims without further evolution or omit the initial reverse step.The three settings were FactArena, “w/o evolution,” and “w/o reverse.”Source: E. Analysis of Arena-driven Claim… · p. 8
  • DeepSeek-R1, Gemini 2.5 Flash, Qwen 3 (235B), and Qwen 3 (32B) had noticeably higher Elo ratings after claim evolution.The paper interprets improved ratings as indicating greater robustness and generalizability to adaptively generated challenging claims.Source: E. Analysis of Arena-driven Claim… · p. 8
  • 100%-accurate simple claims fell to 68% average prediction accuracy after semantic reversal, while later rounds reduced accuracy more slowly and battle counts stabilized at about 220 by rounds 2–3.The paper identifies Reverse as Round 0 and reports that fewer claims remained trivially solvable by all target models as evolution progressed.Source: Figure 6 · p. 8, E. Analysis of Arena-driven Claim… · p. 8
  • Ratings that dropped under evolution indicate greater vulnerability, while drops under FactArena may also suggest potential data leakage in the no-evolution settings.The paper contrasts performance on static benchmarks with performance after claims are transformed into more novel and unseen forms.Source: E. Analysis of Arena-driven Claim… · p. 8

F. Qualitative Comparison with Traditional Benchmark

  • Five human experts compared the same judge agent’s judgments across FactArena and traditional claim-verification benchmarking on 100 randomly sampled battles.They assessed helpfulness, informativeness, soundness, and readability, using the judge agents’ claim-verification guidelines to evaluate the judges themselves.Source: F. Qualitative Comparison with… · p. 8
  • Human experts found FactArena judges’ responses more informative, sound, convincing, and less misleading than responses under traditional claim-verification benchmarking.The blind test sampled 100 battles and asked five experts to compare judgments using helpfulness, informativeness, soundness, and readability criteria.Source: Figure 7 · p. 9, F. Qualitative Comparison with… · p. 9

G. Discussion of Judge Model Version · H. Discussion of Judge Agent Numbers

FactArena’s four-judge panel achieves the highest accuracy among the tested configurations, while using lighter judges preserves strong agreement with more capable models. Adding judges can improve alignment with human preferences, but increases computational and resource costs.

G. Discussion of Judge Model Version

  • Agreement between selected judges and stronger same-family models averages close to 80%, exceeding 90% for DeepSeek-V3 in some cases.The comparison evaluates whether the computationally lighter judge selection affects evaluation reliability.Source: G. Discussion of Judge Model Version · p. 9
  • The selected judge agents were chosen to balance evaluation quality and computational cost rather than relying exclusively on the strongest available LLMs.Each selected judge was compared with a stronger model from the same family.Source: G. Discussion of Judge Model Version · p. 9
  • Using computationally lighter judges does not significantly compromise evaluation quality, and their decisions align with human assessments and stronger same-family LLMs.Source: G. Discussion of Judge Model Version · p. 9

H. Discussion of Judge Agent Numbers

  • The 4-judge configuration achieves the highest accuracy among the tested judge-panel settings, while offering a balance between evaluation quality and efficiency.The authors note that larger panels incur greater computational and resource costs.Source: Table VII · p. 9, H. Discussion of Judge Agent Numbers · p. 9
  • Deepseek-V3 alone reaches 0.67 accuracy, while adding GPT-4o slightly decreases the combined accuracy.The tested sequence adds GPT-4o as the second judge, then Gemini 2.5 Flash as the third, before the four-judge main-experiment panel.Source: H. Discussion of Judge Agent Numbers · p. 9
  • As the panel grows beyond two judges, overall accuracy increases, suggesting that more judge agents can better reflect human preferences.Source: H. Discussion of Judge Agent Numbers · p. 9

I. Analysis of Claim Extraction · J. Case Study · V. CONCLUSION AND FUTURE WORK

FactArena’s stage-wise evaluation, guideline-driven judgment, and adaptive claim evolution reveal model weaknesses that static verification accuracy misses. These sections show how extraction-guideline quality improves, how evolved claims distinguish models, and where the framework could expand.

I. Analysis of Claim Extraction

  • Five experts assessed 20 claims per round, with 0.618 average intra-annotator agreement, using completeness and correctness to evaluate extraction guidelines.Completeness measures coverage of essential information; correctness measures faithful representation without distortion or unnecessary additions.Source: §I · p. 9
  • Guideline quality increased clearly from rounds 0–3, then improved more slowly and converged by rounds 5–6, indicating diminishing returns from further refinement.The evaluation measured completeness and correctness of claim decomposition.Source: §I · p. 9

J. Case Study

  • On the evolved claim, Model A identifies the unsupported seventh sub-claim, while Model B gives redundant analysis and misses the claim’s false relation.The example shows how evolution amplifies performance distinctions and enables more precise assessment of factual understanding.Source: Figure 9 · p. 10, J. Case Study · p. 10
  • Arena-driven evolution analyzes model-specific weaknesses and selectively intensifies claim difficulty, transforming simple verification into a more logically complex test.In the example, both models answer the original and reversed claims correctly, so the judge initially ties them.Source: J. Case Study · pp. 9–10, Figure 9 · p. 10

V. CONCLUSION AND FUTURE WORK

  • Across 16 state-of-the-art models, FactArena produced robust and fair rankings, improved inter-judge consistency, and exposed performance discrepancies that static accuracy metrics miss.Its adaptive claim evolution also probed robustness and revealed reasoning limitations hidden by fixed test sets.Source: §V · p. 10
  • FactArena integrates stage-wise auditing, guideline-driven judgment, and arena-based claim evolution to assess LLM factual reasoning more comprehensively than verification-only evaluation.The authors present it as a foundation for rigorous, scalable, human-aligned evaluation in safety-critical factual reasoning.Source: §V · p. 10
  • Future work could add multimodal claims,, richer retrieval sources, more diverse judges, uncertainty-aware evaluation, and real-world testing on noisy, adversarial, or rapidly evolving claims.The suggested retrieval sources include structured databases and temporal information.Source: §V · p. 10

APPENDIX A PROMPT TEMPLATES

Appendix A provides curated prompt templates for claim extraction, evidence retrieval, justification and verdict prediction, claim evolution, guideline consolidation, and stage-wise judgment.

Claim Extraction Prompt

  • Claim extraction instructs target models to break claims into sub-claims of proper granularity.The prompt is shown in Figure 10.Source: Appendix A · p. 13

Evidence Retrieval Prompt

  • Evidence retrieval prompts models to generate factual evidence from the claim, its sub-claims, and web information retrieved with an external tool.The prompt is shown in Figure 11.Source: Appendix A · p. 13

Justification Production & Verdict Prediction Prompt

  • Justification and verdict prediction prompts models to use previously generated sub-claims and evidence to justify and give a final claim verdict.The prompt is shown in Figure 12.Source: Appendix A · p. 13

Evolution Prompt

  • Claim evolution reverses semantics into a logical, verifiable alternative with the opposite veracity label, or targets model weaknesses with more challenging claims.The evolver avoids superficial or trivial modifications when reversing semantics.Source: Appendix A · p. 13

Claim Extraction Guideline Consolidation Prompt

  • Claim extraction guideline consolidation blinds judges to sub-claim sources and randomly shuffles model-answer order to mitigate potential bias.The consolidation prompt is shown in Figure 15.Source: Appendix A · p. 13

Arena-styled Stage-wise Judgment Prompt

  • Stage-wise judgment randomly orders model answers to avoid positional bias and compares target-model performance at each claim-verification stage.The judgment prompt is shown in Figure 16.Source: Appendix A · p. 13

APPENDIX B DETAILED CASES

Appendix B illustrates FactArena’s pipeline, guideline consolidation, and claim evolution through examples. The cases show how claim evolution increases semantic complexity and verification difficulty.

A Case of Fact-checking Pipeline

  • FactArena’s three-stage pipeline decomposes a claim into sub-claims, retrieves supporting web evidence, and combines these with the original claim to predict a verdict.The stages are claim extraction, evidence retrieval, and claim verification.Source: Appendix B · p. 13, Figure 17 · p. 15

A Case of Guideline Consolidation

  • The consolidated claim-extraction guideline retains the reference decomposition strategy while adopting model-answer expression style, yielding a more coherent and verifiable guideline.The appendix illustrates this synthesis with an example.Source: Appendix B · p. 14, Figure 18 · p. 16

More Cases of Claim Evolution

  • Across two rounds of arena-driven evolution, claims become semantically more complex, making verification more challenging.The cases show original, semantically reversed, and evolved claims.Source: Appendix B · p. 14, Figure 19 · p. 16

Detailed Case of Arena-driven Claim Evolution

Related papers

Loading 2601.02669v1…