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

Existing LLM fact-checking benchmarks largely assess claim verification while omitting claim extraction and evidence retrieval. FactArena addresses this gap with an automated, arena-style framework covering all pipeline stages, collective judgment, and adaptive claim evolution. Across 16 models, it produces robust rankings and exposes discrepancies that static accuracy metrics miss.

  • Problem

    Existing evaluations primarily focus on claim verification rather than the full fact-checking workflow, limiting assessment of stage-wise factual reasoning and the constraints of fixed test sets.

  • Method

    FactArena combines automated stage-wise fact-checking, arena-based collective judgment with Elo and Bradley–Terry ranking, and adaptive evolution of challenging claims.

  • Results

    Across 16 state-of-the-art models, FactArena produces robust and fair rankings, reveals performance discrepancies missed by static accuracy metrics, and probes deeper reasoning limitations beyond fixed test sets.

  • Takeaways & Limitations

    Holistic, automated benchmarking offers a broader basis for diagnosing LLM factual reasoning than verification-only evaluation.

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 addresses the limits of verification-only evaluation by benchmarking LLMs across the full fact-checking pipeline. Its automated arena framework combines stage-wise assessment, collective judgments, and adaptive claim evolution to support broader auditing of factual reasoning and robustness.

  • Motivation: Fact-checking comprises claim extraction, evidence retrieval, and claim verification, but traditional LLM evaluations primarily assess only verification.The overlooked preliminary stages identify claims and gather supporting or refuting sources.
  • Motivation: Comprehensive assessment across every stage is essential because LLMs are increasingly deployed throughout the fact-checking pipeline, including for complex real-world claims.The paper links full-pipeline auditing to deeper insight into model limitations and reliable deployment.
  • Framework: FactArena is a systematic, automated arena framework that evaluates LLMs across the entire fact-checking pipeline using unbiased collective judgments.Its design emphasizes holistic coverage, diverse evaluator perspectives, and scalable automation.
  • Framework: The framework conducts peer-battle arenas for claim decomposition, tool-augmented evidence retrieval, and explanation and verdict quality.These stages align the evaluation process with the operational stages of fact-checking.
  • Framework: FactArena adaptively generates increasingly challenging and unseen claims to probe LLM factual robustness and knowledge boundaries beyond fixed test sets.Claim evolution provides a dynamic complement to static evaluation data.
  • Findings: Extensive experiments report improved reliability and completeness of workflow evaluation, with diverse agent judgments showing strong consistency and close alignment with human experts.The reported judgments are intended to provide actionable information for trustworthy LLM auditing.

II. RELATED WORK

FactArena addresses limitations in prior fact-checking evaluations by benchmarking LLMs across the complete pipeline rather than claim verification alone. It operationalizes this pipeline through claim extraction, evidence retrieval, and justification-based verdict prediction.

  • Prior fact-checking evaluations primarily assess claim verification, despite fact-checking also requiring claim extraction and evidence retrieval.
  • FactArena frames fact-checking as decomposing complex claims, retrieving external evidence, and producing justified verdicts.
  • Claim extraction converts a complex claim into verifiable sub-claims that guide subsequent verification planning.
  • Evidence retrieval uses Google Search and summarizes retrieved web information as evidence for or against the original claim.
  • Claim verification processes the original claim, sub-claims, and evidence to generate a justification and predicted verdict.

C. Arena-styled Stage-wise Judgment

FactArena compares anonymous models through stage-wise peer battles guided by consolidated reference standards. It converts evaluator judgments into rankings using Elo and Bradley–Terry models, with the latter reducing order sensitivity.

  • FactArena conducts anonymous peer battles across claim extraction, evidence retrieval, and claim verification outputs.
  • Claim extraction: Judges consolidate diverse sub-claim decompositions into a shared guideline and rotate judges and samples until all viewpoints are incorporated.
  • Evidence retrieval: Evidence retrieval is judged against factual reference information from Wikipedia, which serves as a basis rather than a golden label.
  • Claim verification: Verification justifications are assessed for helpfulness, informativeness, soundness, and readability alongside verdict prediction.
  • Elo ratings summarize head-to-head outcomes, while Bradley–Terry estimates relative strengths without depending on match-up order.

D. Arena-driven Claim Evolution

FactArena evolves claims to probe model limitations beyond fixed test sets. When models answer correctly, it reverses claim semantics and then generates increasingly challenging, semantically equivalent variants using arena feedback.

  • Fixed fact-checking test sets may provide superficial assessments and risk test leakage, limiting discovery of models’ deeper weaknesses and knowledge boundaries.
  • When all target models predict a claim’s factuality correctly, an evolver first creates a contrastive claim with the opposite verdict.
  • If models also correctly predict the contrastive claim, arena judgments drive iterative generation of increasingly challenging, semantically equivalent variations.
  • Each evolved claim is judged and stored in a record pool, enabling identification of scenarios where models perform poorly.

IV. EXPERIMENTS

Experiments evaluate FactArena on multi-hop claims using a broad set of models and repeated pairwise comparisons. The setup covers 16 LLMs from seven families and produces thousands of judgments for ranking.

  • Experimental settings: The experiments use complex claims from HOVER and FEVEROUS that require multi-hop reasoning beyond simple factual matching.
  • Reported outcomes: Table II reports model rankings across multiple evaluation dimensions, including battle count and accuracy percentage, sorted by overall judge.
  • Models: 16 LLMs across 7 model families are evaluated, covering Claude, DeepSeek, GPT, Gemini, Grok, Llama, and Qwen models.
  • Data and comparisons: Each of 400 claims receives 8 target-model responses per task for pairwise comparisons.
  • Data and comparisons: Approximately 13,000 valid judgments are collected, with each target LLM participating in about 1,600 comparisons.

C. Main Results

FactArena yields stage-wise rankings that expose model strengths and weaknesses beyond verification-only accuracy, while reference guidelines improve judge consistency and fairness. Human-aligned evaluations further support the reliability of its judgments.

  • GPT-o3 and DeepSeek-R1 rank first and second overall by Elo, consistently outperforming other models across all fact-checking stages.
  • Gemini 2.5 Pro surpasses GPT-4.5 in claim extraction and verification but is weaker in evidence retrieval, demonstrating distinct stage-wise capabilities.
  • Static claim-verification accuracy can overstate competence, as Gemini 2.5 Pro’s advantage diminishes under full-pipeline evaluation.
  • Judgment reliability: FactArena produces more robust and stratified rankings than its no-guideline setting, while guidelines reduce judge-specific and same-family preferences.
  • Judgment reliability: Evaluation guidelines generally improve inter-judge agreement, particularly for claim extraction and justification production.
  • Human alignment: Human experts provide a comparison basis for judge-agent consistency, with reported inter-annotator agreement of 0.632 and intra-annotator agreement of 0.726.

E. Analysis of Arena-driven Claim Evolution

Arena-driven claim evolution adaptively increases evaluation difficulty and reveals robustness differences that static or trivially transformed claims may miss. Its effects vary across models, exposing both generalization and vulnerability patterns.

  • DeepSeek-R1, Gemini 2.5 Flash, Qwen 3 (235B), and Qwen 3 (32B) gain Elo ratings after claim evolution, indicating stronger robustness to generated challenges.
  • Models whose ratings decline under evolution exhibit greater vulnerability, while some declines suggest static settings may permit potential data leakage.
  • Semantic reversal reduces average prediction accuracy from 100% to 68%, making initially simple claims substantially more challenging.
  • Blind comparison: FactArena judges produce more informative, sound, convincing, and less misleading responses than traditional verification-only benchmarking in a human blind test.

G. Discussion of Judge Model Version

FactArena examines judge-model strength, panel size, and guideline quality as components of evaluation reliability. Lighter judges remain aligned with stronger models, while larger panels improve human alignment at added cost.

  • Judge model version: These results indicate that computationally lighter judges do not significantly compromise evaluation quality and remain aligned with human assessments and stronger models.
  • Judge agent numbers: The four-judge configuration achieves the highest accuracy among compared panel sizes, although adding judges increases computational and resource costs.
  • Judge agent numbers: DeepSeek-V3 alone reaches accuracy 0.67, while adding GPT-4o slightly lowers combined accuracy before larger panels improve alignment.
  • Guideline quality: Claim-extraction guideline quality increases from rounds 0 to 3 according to completeness and correctness evaluations.

J. Case Study

The case study illustrates how arena-driven evolution transforms easy verification examples into logically complex claims that expose model-specific weaknesses. This supports FactArena’s broader finding that adaptive claims reveal limitations hidden by static accuracy.

  • The original and reversed Schmeichel claims are correctly answered by both models, producing a tie judgment.
  • Arena-driven analysis identifies Model A’s less structured reasoning and Model B’s unnecessary detail as distinct weaknesses for claim evolution.
  • On the evolved claim, Model A detects an unsupported seventh sub-claim, whereas Model B misses the false relation amid redundant information.
  • Adaptive claim evolution amplifies performance distinctions and exposes reasoning limitations that fixed test sets can leave hidden.
  • FactArena’s broader evaluation over 16 models produces fair rankings and reveals discrepancies that static accuracy metrics fail to capture.

APPENDIX A PROMPT TEMPLATES

Appendix A presents the prompt templates that operationalize FactArena’s fact-checking stages, judgment process, and arena-driven claim evolution. The templates standardize model outputs and reduce positional or source-related bias in evaluation.

  • Prompt Templates: Claim extraction instructs target models to decompose claims into sub-claims of appropriate granularity.
  • Prompt Templates: Evidence retrieval combines the claim, sub-claims, and externally retrieved web information to produce factual evidence.
  • Prompt Templates: Justification and verdict prediction use the generated sub-claims and evidence to produce a rationale and final claim verdict.
  • Prompt Templates: Claim evolution reverses claim semantics or generates harder claims while preserving logical verifiability and targeting model weaknesses.
  • Prompt Templates: Guideline consolidation blinds judges to sub-claim sources and randomly shuffles model-answer order to mitigate bias.
  • Prompt Templates: Stage-wise judgment randomly orders model answers and compares target-model performance at each claim-verification stage.

APPENDIX B DETAILED CASES

Appendix B illustrates FactArena’s three-stage pipeline, guideline consolidation, and adaptive claim evolution through worked examples. The cases show how arena judgments identify reasoning weaknesses and guide more challenging, semantically controlled claims.

  • A Case of Fact-checking Pipeline: FactArena’s case pipeline decomposes a claim, retrieves supporting web information, and integrates the evidence during claim verification.
  • A Case of Guideline Consolidation: Guideline consolidation preserves the reference decomposition strategy while incorporating model-answer expression styles into a coherent, verifiable guideline.
  • More Cases of Claim Evolution: Claim evolution progressively increases semantic complexity across reversed and arena-evolved claims, making verification more challenging.
  • Detailed Case of Arena-driven Claim Evolution: The evolution prompt asks agents to identify model weaknesses and reformulate claims into logically equivalent or semantically faithful versions with greater reasoning complexity.
  • A Case of Fact-checking Pipeline: The Denton example demonstrates a claim labeled Refutes within the FactArena fact-checking pipeline.
  • Detailed Case of Arena-driven Claim Evolution: The reversed Schmeichel claim changes the semantic direction while retaining a verifiable relation involving football recognition.
  • A Case of Fact-checking Pipeline: The extraction examples contrast a detailed decomposition naming Peter Schmeichel with a more compressed decomposition referring to the father generically.
  • A Case of Fact-checking Pipeline: The retrieval examples report that Peter Schmeichel received the IFFHS World’s Best Goalkeeper recognition in both 1992 and 1993, contradicting the reversed claim.
Loading 2601.02669v1…