PaperlayerView exact revision on arXiv

Source-linked AI summary

How Robust Are Automated Fact-Checking Systems? A Cross-Benchmark Evaluation

Aida Usmanova, Zangir Iklassov, Markus Leippold, Ricardo Usbeck

arXiv:2608.25934v1cs.AIcs.LG

TL;DR

Automated fact-checking evaluations often omit simple baselines and test systems on only one benchmark, leaving cross-domain generalisation and full-pipeline robustness uncertain. This paper evaluates nine models across four diverse datasets and finds domain-dependent rankings, competitive classical baselines, and retrieval as the primary bottleneck. It also documents scope limits involving dataset coverage, model coverage, and prompt sensitivity.

  • Problem

    The paper addresses the lack of cross-domain evaluation of complete retrieve-then-verify AFC pipelines with simple baselines.

  • Method

    It benchmarks nine models across four datasets, covering random and sparse baselines, fine-tuned transformers, zero-shot LLMs, and top AVeriTeC 2025 systems.

  • Results

    System rankings depend strongly on domain and metric, classical baselines can outperform evidence-conditioned systems, and retrieval remains the primary bottleneck.

  • Takeaways & Limitations

    The findings support evaluating AFC systems across multiple domains and metrics rather than treating single-benchmark rank as general capability.

  • Takeaways & Limitations

    Conclusions are limited to four datasets and do not cover multilingual or multi-domain AFC, recent generative systems beyond Llama 3.1, or prompt sensitivity.

Abstract

from arXiv · show

Automated fact-checking (AFC) systems retrieve evidence and predict claim veracity, yet evaluations omit simple baselines, systems are developed for a single benchmark and cannot be trusted to generalise across domains. No prior work cross-evaluates the full two-stage retrieve-then-verify pipeline across diverse datasets, complementing retrieval-only studies (Thakur et al., 2021) and single-stage benchmarking studies (Calamai et al., 2025). We benchmark nine models, ranging from random and sparse baselines to fine-tuned transformers, zero-shot LLMs, and the two highest-ranked systems from the AVeriTeC 2025 shared task, across four datasets spanning scientific, open-web, and climate domains. Three findings stand out: (1) on ClimateCheck claim-only and fine-tuned models outperform zero-shot LLM and top-performing AVeriTeC 2025 systems, highlighting that noisy evidence can degrade veracity prediction; (2) system rankings are strongly domain- and metric-dependent: the best model on SciFact (macro-F1 0.70) drops to 0.31 on ClimateCheck, while the AVeriTeC 2025 winner and runner-up swap rankings based on evaluation metrics and datasets; (3) replacing retrieved evidence with gold annotations improves veracity accuracy by 14-22 points across models, confirming retrieval remains primary bottleneck. We release code, pre-processed datasets, and all results to support reproducible AFC research.

1 Introduction

The paper asks whether automated fact-checking progress generalises beyond individual benchmarks and evaluates the complete retrieve-then-verify pipeline across diverse datasets. Its results show strong baseline competitiveness, domain- and metric-dependent rankings, and retrieval as the main bottleneck.

  • Motivation: AFC systems may not generalise across domains because they are usually developed and tested on single benchmarks without simple lower-bound baselines.The paper argues that absent baselines obscure whether gains reflect language understanding or dataset-specific engineering.
  • Research gap: Prior studies examined retrieval or single-task classification separately, but no prior work cross-evaluated the complete retrieve-then-verify pipeline across structurally diverse domains.
  • Approach: The study evaluates four diverse datasets under identical conditions using sparse retrieval, fine-tuned transformers, zero-shot LLMs, and leading AVeriTeC 2025 systems.
  • Findings: Logistic regression over sparse retrieval outperformed evidence-conditioned LLMs and shared-task top performers because harmful evidence retrieval degraded verification.
  • Findings: 14-22 points: replacing retrieved evidence with gold annotations improved veracity accuracy across models, showing retrieval quality is a main challenge.
  • Findings: Macro-F1 0.700 on SciFact fell to 0.315 on ClimateCheck, while AIC CTU and SANCTUARY swapped rankings across datasets and metrics.

2 Related Work

Related work established automated fact-checking pipelines, benchmark datasets, retrieval methods, and reproducibility concerns. However, prior benchmarking largely isolated pipeline stages rather than evaluating complete AFC systems across diverse domains.

  • Automated fact-checking: AFC veracity prediction is commonly framed as natural language inference over retrieved evidence, with extensions for graph aggregation, explanation generation, and contrastive robustness.
  • Automated fact-checking: Recent agentic systems such as FIRE and DEFAME let the verifier control additional evidence retrieval, moving beyond fixed retrieve-then-verify pipelines.
  • Benchmarks: Fact-checking benchmarks span political and open-web claims, scientific claims against biomedical abstracts, climate claims linked to Wikipedia, and live-web evidence retrieval.
  • Evidence retrieval: Sparse, dense, and retrieval-augmented methods have advanced evidence retrieval, but specialised training remains necessary for domain-specific AFC corpora.
  • Benchmarking studies: Reproducibility studies found classical baselines often remain close to fine-tuned models, while heterogeneous retrieval benchmarks frequently favor BM25 over dense models.

3 Tasks, Datasets, and Models

The study evaluates a two-stage AFC pipeline across four datasets organized by evidence domain and claim origin, comparing simple, neural, LLM, and shared-task systems under common metrics.

  • Tasks: Given claim c and corpus D, the pipeline retrieves K relevant documents and predicts Supports, Refutes, or NEI from the claim and retrieved evidence.
  • Datasets: The four datasets cover open-web AVeriTeC, Wikipedia-based ClimateFEVER, expert-written SciFact, and social-media ClimateCheck claims.
  • Datasets: Dataset structure varies along evidence domain and claim origin, with scientific evidence requiring semantic matching and informal claims widening vocabulary gaps.
  • Models: The evaluation uses random and sparse retrieval, fine-tuned transformers, zero-shot LLMs, and AVeriTeC 2025 winner and runner-up systems.
  • Models: Claim-only baselines diagnose whether labels can be predicted from claim text without evidence, probing dataset artifacts or annotation shortcuts.
  • Evaluation: Retrieval is measured with Recall@K and F1@K, while veracity is measured using accuracy and equally weighted macro-averaged F1.

4 Experimental Results and Analysis

Results across four datasets show unstable system rankings, strong domain effects, and retrieval quality as the main constraint on veracity prediction. Classical and fine-tuned models can outperform larger evidence-conditioned systems when retrieved evidence is noisy or mismatched.

  • Evidence retrieval: Retrieval rankings reverse by domain: AIC CTU reaches R@5 = 0.731 on SciFact versus TF-IDF’s 0.217, while TF-IDF leads AVeriTeC at R@5 = 0.126.TF-IDF consistently outperforms BM25, while ClimateCheck and ClimateFEVER remain difficult for all methods.
  • Veracity prediction: On ClimateCheck, LONGFORMER achieves the highest accuracy at 0.618, narrowly followed by DISTILROBERTA at 0.617.Retrieved evidence can reduce LLM accuracy under domain shift, and low-quality retrieval can mislead veracity prediction.
  • Veracity prediction: On SciFact, SANCTUARY leads with accuracy 0.702 and macro-F1 0.700, while fine-tuned models remain below 0.47 accuracy.SANCTUARY’s sentence-level evidence grouping aligns with SciFact’s rationale-sentence annotations.
  • Claim-only signals: Claim-only macro-F1 exceeds random prediction by 26.7 points on ClimateCheck, indicating substantial label-predictive information in claim text alone.Weak evidence annotations and annotation ambiguity may contribute to this apparent baseline advantage.
  • Retrieval quality: 14–22 percentage points: gold evidence improves veracity accuracy across LLM models, confirming retrieval as the primary bottleneck.The largest gains occur for LLAMA 70B (+22 pp) and LLAMA 8B (+20 pp) on ClimateFEVER.
  • Cross-dataset rankings: No single system dominates across benchmarks: SANCTUARY reaches macro-F1 0.700 on SciFact but 0.315 on ClimateCheck.AIC CTU and SANCTUARY also swap rankings depending on evaluation metrics and datasets.

5 Conclusion

The study shows that simple baselines remain necessary, while AFC performance varies across domains and metrics and retrieval quality remains central to system performance.

  • The benchmark evaluates evidence retrieval and claim veracity prediction across AVeriTeC, SciFact, ClimateCheck, and ClimateFEVER.
  • Claim-only TF-IDF + LOGREG outperform evidence-conditioned zero-shot LLMs and top-performing systems on ClimateCheck.Misleading retrieval can substantially degrade AFC systems.
  • No single system dominates across all four benchmarks, with rankings changing by domain and evaluation metric.SANCTUARY spans 0.39 macro-F1 across datasets, while AIC CTU and SANCTUARY swap positions depending on the evaluation setting.
  • Replacing retrieved evidence with gold annotations improves accuracy across models and datasets, identifying retrieval as the primary bottleneck.
  • Cross-domain AFC evaluation should include classical baselines and explicitly separate retrieval from veracity prediction.Leaderboard rankings should not be treated as proof of real-world utility until systems demonstrate reliability across diverse domains.

Limitations

The conclusions are bounded by the study’s four-dataset scope and by limitations in its failure analysis.

  • The four-dataset study may not generalise to all AFC domains or claim types.FEVER and FEVERous were excluded because their scale exceeded the computational budget; multilingual, multi-domain, and newer generative systems were also not evaluated.
  • Failure analysis used two annotators with moderate agreement and covered only a subset of error cases.

Ethical Considerations

The study uses publicly released benchmarks and existing models rather than collecting new data or deploying a production fact-checking system.

  • The evaluation uses AVeriTeC, SciFact, ClimateFEVER, and ClimateCheck, with no newly collected data.Claims and evidence documents come from the original datasets, whose licences permit academic use.

Generative AI Usage

Generative AI tools were used only to improve the manuscript’s grammar, typos, readability, and written quality.

  • ChatGPT was used to check grammar and typos and enhance the readability and quality of the written text.

A Evaluation Metrics

The evaluation defines retrieval and veracity metrics for comparing evidence selection and classification under dataset-specific conditions. It also highlights class imbalance, retrieval bias, and annotation ambiguity as important interpretive factors.

  • Retrieval metrics: Recall@K measures coverage of gold evidence, Precision@K measures retrieved-result quality, and F1@K combines both through their harmonic mean.
  • Retrieval metrics: Recall@K plateaus when K exceeds the gold-set size, remaining constant across K ∈{5, 10, 20} for AVeriTeC because its median gold set has one document.
  • Retrieval metrics: BM25’s length normalisation penalises long documents, explaining why it underperforms TF-IDF on long Wikipedia and scientific-abstract corpora.
  • Veracity metrics: Macro F1 weights classes equally, unlike accuracy, and can differ from accuracy by up to 20 points when the majority class reaches 60% of samples.
  • Metric choice: Recall@K and F1@K are preferred to MRR when claims require several gold documents, while rank-based metrics preserve the same sparse-retriever ordering.
  • Dataset and result interpretation: Class imbalance drives stronger performance on dominant labels and near-zero performance on rare labels, especially Conflicting Evidence, while ClimateCheck NEI is learned more reliably by transformers.
  • Dataset and result interpretation: Moderate agreement and annotation judgments complicate small score differences: annotators agreed at κ = 0.55 and classified 38% of sampled errors as annotation mistakes or debatable cases.
  • Dataset and result interpretation: Longer documents are retrieved more often on SciFact and ClimateCheck because claim-adjacent tokens inflate Recall@K independently of true relevance.

E.2 Per-Dataset Analysis

Per-dataset errors arise from evidence fragmentation, lexical mismatch, multi-passage requirements, label ambiguity, and dataset-specific annotation schemes. Cross-dataset results show that Conflicting Evidence is broadly difficult and annotation issues limit interpretation of small margins.

  • AVeriTeC: AVeriTeC retrieval often returns evidence fragments that are individually insufficient to support or refute claims.
  • AVeriTeC: SANCTUARY over-predicts Refuted on AVeriTeC, while Conflicting Evidence remains nearly unpredictable with F1 < 0.18 for all models.
  • SciFact: SciFact failures combine lexical mismatch with sentence-level rationale granularity, causing retrievers to miss semantically related evidence and verifiers to predict NEI.
  • ClimateCheck: ClimateCheck has the widest vocabulary gap: informal social-media claims and formal scientific abstracts frequently share few tokens, producing irrelevant retrieval.
  • ClimateFEVER: ClimateFEVER typically requires 2–5 Wikipedia passages, so single-passage retrieval can be insufficient; this yields the study’s largest oracle gains of +20–22 pp.
  • Cross-dataset patterns: Across datasets, Conflicting Evidence scores 0.00–0.20, LLMs often over-predict NEI, and annotation issues make margins below 5 accuracy points unreliable.

F Frontier-LLM Verifier Experiments

Frontier-LLM experiments isolate evidence quality by holding the verifier and prompt fixed, then compare claim-only, retrieved, gold, and iterative retrieval conditions. Gold evidence performs best, while retrieval quality constrains both one-shot and agentic systems.

  • Experimental design: The experiments hold CLAUDE OPUS 4.8 and its prompt fixed while varying evidence conditions on fixed AVeriTeC and ClimateCheck samples.
  • Three evidence conditions: Gold evidence produces the best AVeriTeC accuracy at 87.3, compared with 81.0 for claim-only input and 38.1 for retrieved evidence.
  • Three evidence conditions: A 49-point AVeriTeC gap between gold and retrieved evidence is attributable to evidence quality alone under the fixed-model comparison.
  • Iterative retrieval: Agentic retrieval remains below oracle performance—87.3 on AVeriTeC and 70.9 on ClimateCheck—with second rounds triggered for 25.4% and 30.9% of claims.

G Recommendations for AFC Evaluation Design

The paper recommends evaluation practices that expose baseline competitiveness, domain transfer, retrieval limits, class-specific behavior, and annotation quality. These practices are intended to make AFC comparisons more reliable and interpretable.

  • Baseline and scope: TF-IDF and BM25 with logistic regression should be mandatory starting points for assessing dataset difficulty and improvement beyond surface-level matching.
  • Baseline and scope: A proposed system should beat TF-IDF + LOGREG by more than 5 accuracy points in-domain before its improvement is considered reliable amid annotation noise.
  • Baseline and scope: Single-benchmark results are insufficient: cross-dataset evaluation should include at least one dataset outside the system’s training distribution.
  • Pipeline diagnosis: Oracle-retrieval experiments should separate retrieval and veracity errors by revealing the upper bound available from better evidence.
  • Metric reporting: Per-label F1 should accompany aggregate metrics because macro F1 and accuracy can differ by up to 20 points under class imbalance.
  • Annotation quality: Annotation quality should be quantified through inter-annotator agreement or manual sampling before claiming improvements smaller than typical noise margins.
Loading 2608.25934v1…