Evaluating and Improving Evidence-Grounded Fact-Checking in LLMs via Multi-Round Evidence Ablation
Xingyu Deng, Mingzi Cao, Nikolaos Aletras, Xi Wang, Mark Stevenson
cs.CL · cs.AI · cs.IR
Quick verdict
FAE ablation evaluation shows LLM verifiers persist sans cited evidence; REAL training fixes it, but single-seed runs and one table inconsistency temper trust.
Summary
The paper asks whether LLM fact-checkers actually use the evidence they cite or quietly fall back on parametric knowledge. It introduces FAE (Fact-Ablated Evaluation), an iterative counterfactual protocol: after an initial prediction, the model's own cited evidence sentences are removed round by round and the claim is re-verified, with three metrics (Immediate Sensitivity, End-State Retention, Ideal Offset) scoring how quickly predictions collapse to NOT_ENOUGH_INFO. Off-the-shelf LLMs (GPT-4o-mini, Gemini-2.5-Flash-Lite, Qwen-2.5-32B, Llama-3.1-8B) keep high accuracy under ablation, indicating knowledge-driven decisions. The proposed REAL fine-tunes verifiers with paired supervision — full context yields label plus evidence; gold-evidence-ablated context yields abstention — plus cross-model evidence enhancement (majority vote of GPT-4o-mini, Gemini-2.5-Flash-Lite, Qwen-2.5-32B) to patch incomplete annotations. With LoRA on FEVER, REAL reaches 95.66 label accuracy and IS 99.06 (base Llama: 43.79), stays ahead out-of-domain (SciFact 89.39, Climate-FEVER 74.86, Check-COVID 88.90), and ablations attribute grounding to the counterfactual component. Representation analysis shows base models drift toward knowledge-based states during ablation while REAL-trained ones stay flat. Limitations: single fixed seed with no significance tests, a Check-COVID table inconsistency between Tables 1 and 2, exclusion of gold-NEI claims, and residual dependence on possibly incomplete gold annotations.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Off-the-shelf LLM verifiers keep their SUPPORT/REFUTE predictions even after the evidence they cited is iteratively ablated, revealing reliance on parametric knowledge (§4.2, Fig. 3).
- FAE operationalises evidence dependency via three metrics — Immediate Sensitivity, End-State Retention and Ideal Offset — computed over the ablation trajectory (§4.3).
- REAL's paired counterfactual supervision (full context → label+evidence; ablated context → NEI) yields near-ideal ablation behaviour (FEVER IS 99.06 vs 43.79 for base Llama) while improving accuracy over off-the-shelf models (Table 1).
- Cross-model majority-vote evidence enhancement is needed to prevent corrupted counterfactual supervision from incomplete annotations, mainly protecting evidence-selection quality (§5.2, Table 2).
- During ablation, base verifiers' hidden states drift toward their knowledge-based state, whereas REAL-trained verifiers stay flat, indicating suppression of the parametric fallback (§7.4, Fig. 6).
Strengths
Dynamic counterfactual evaluation of the model's own citations
FAE (§4.1) iteratively removes the verifier's own predicted evidence rather than gold evidence, going beyond the static gold-removal of Akhtar et al. [2]; the three metrics IS/ER/IO (§4.3) make evidence dependency a measurable behavioural quantity.
Clean component attribution in the ablation study
Table 2 shows Standard SFT reaches 96.18 label accuracy on FEVER yet collapses to IS 10.75 / ER 11.02, while w/o Evidence Enhancement keeps IS 98.82 — isolating the counterfactual ablation term as the driver of grounding and the enhancement step as the protector of evidence quality (Climate-FEVER F1 41.26 vs 66.68).
Out-of-domain transfer mitigates the shortcut concern
Honest handling of annotation incompleteness
Behavioural artifact ruled out
Weaknesses
Internal numerical inconsistency
Thin statistical reporting
Accuracy trade-off versus the proper baseline is underplayed
No external training-side baselines
Confound between evidence selection and parametric reliance
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets a genuine blind spot in LLM-based fact-checking: label accuracy alone cannot distinguish a verifier that reasons over retrieved evidence from one that recovers memorised facts. Three claims are made: (1) off-the-shelf LLMs keep predicting SUPPORT/REFUTE even after the evidence they themselves cited is iteratively removed, revealing parametric-knowledge reliance; (2) FAE, a multi-round counterfactual protocol with three metrics (IS, ER, IO), exposes this behaviour; (3) REAL, counterfactually supervised fine-tuning plus cross-model evidence enhancement, restores evidence dependency. The scope is deliberately narrow: verification only (documents provided, no retrieval), gold-NEI claims excluded, four datasets (FEVER in-domain; SciFact, Climate-FEVER, Check-COVID out-of-domain), LoRA fine-tuning of Llama-3.1-8B-Instruct with a Qwen-2.5-7B replication.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is FAE (Fact-Ablated Evaluation)?
FAE is a counterfactual test for LLM fact-checkers. The model verifies a claim and cites evidence sentences; those cited sentences are removed and the model re-verifies the same claim, round after round. A truly evidence-grounded model should switch to NOT_ENOUGH_INFO (abstain) once its cited evidence is gone. Three summary metrics — Immediate Sensitivity, End-State Retention and Ideal Offset — score how quickly and completely that collapse happens.
What does the REAL training framework do?
REAL fine-tunes a verifier on paired examples: with full context, the target output is the correct veracity label plus evidence; with the gold evidence deliberately removed from the context, the target output is abstention (NOT_ENOUGH_INFO). This makes predictions causally dependent on evidence availability. A pre-processing step uses three LLMs voting on which sentences support each claim to fix incomplete evidence annotations before building the ablated inputs.
Did REAL hurt fact-checking accuracy?
Compared with off-the-shelf LLMs (GPT-4o-mini, Gemini-2.5-Flash-Lite, Qwen-2.5-32B, Llama-3.1-8B), no — REAL's accuracy is higher on all four benchmarks (e.g., 95.66 label accuracy on FEVER vs 87.24 for GPT-4o-mini). Compared with standard supervised fine-tuning, however, REAL gives up some accuracy — about 10 label-accuracy points on Climate-FEVER (74.86 vs 84.45) and 1-2 points on FEVER and SciFact — in exchange for near-perfect evidence sensitivity.
Is code and data available for this paper?
The paper links a public GitHub repository (github.com/xingyu-deng/FAE_REAL) covering data preparation and experiments, and all four datasets (FEVER, SciFact, Climate-FEVER, Check-COVID) are public. Reproduction needs GPU fine-tuning of an 8B model with LoRA plus API access to the three models used for evidence enhancement, so results should be treated as medium-ease to replicate.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 3-6 appear in the extraction only as axis fragments and isolated values (e.g., Fig. 4's 13.3%/13.4%/73.3%); curve shapes cannot be independently verified from text, so trajectory-level judgments rest on the authors' descriptions and the tabulated IS/ER/IO values.
- No truncation was detected: the text ends at reference [78], consistent with the in-text citation range, and all four tables are present, though flattened table headers mean column mapping was reconstructed from context.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.