Papers for

fact-checking teams

Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.

Model citations often come from memory not retrieved sources

Cited but Unused: Parametric Provenance in Retrieval-Augmented Generation

Abstract: Retrieval-augmented generation systems attach citations to attribute claims to their sources. These citations frequently misstate provenance, and standard verification cannot detect this. We measure provenance by counterfactual value substitution: a fact in the retrieved passages is replaced with a false variant and the model re-queried. We find that across five sizes of one model family and a second comparator family, 1.2% to 15.7% of citable claims derive from parametric memory rather than the retrieved context. At the most conservative substitution distance these figures are 4.6% and 8.5% on the two models we tested at it, with the rate rising with model size within a family. We further find that natural language inference verification, the standard approach, gives results indistinguishable from chance across models, even when restricted to the cited passage. We determine that the cause is structural: a model answering from memory selects a citation that agrees with its output. Comparing ten attribution methods on identical claims, the contrast between the model's distribution with and without context best separates provenance. Acting on that signal does not improve provenance honesty at these base rates. The resulting detector is hence still not accurate enough to act on: it measures what the model knew rather than what it used, and separating those remains the open problem.

Thu 12 NovArtificial Intelligence
The gist
Sometimes when AI systems generate answers with citations, the cited documents aren't actually the source of the information. The authors tested this by swapping true facts in the retrieved documents with false ones and checking if the model’s answers changed. They found that in many cases, the model relied on what it had memorized instead of the cited passages. Existing methods to verify citations often fail to detect this problem because the model picks citations that match its answer, whether or not they were actually used. The authors created a better way to detect true provenance but it is still not reliable enough to trust completely.
Open → 2611.17161v1

Efficient method improves fact checking of claims from documents

Label-Efficient Document-Grounded Factuality Detection via Cross-Paraphrastic Invariance Learning

Abstract: Large language models (LLMs) can generate fluent claims that are unsupported by a given source document, motivating accurate and efficient document-grounded factuality detection. Existing LLM-based evaluators often require repeated inference and incur high computational cost, whereas compact factuality detectors are efficient at inference time but commonly rely on large labeled or synthetically generated corpora. We propose CPIL (Cross-Paraphrastic Invariance Learning), a two-stage Siamese framework that converts a small labeled set into relational supervision without acquiring additional factuality labels. CPIL constructs positive pairs from distinct label-preserving paraphrastic views of the same document-claim instance and negative pairs from opposite-label examples associated with the same source document. It further introduces quantile-aware semi-hard negative mining, which derives a boundary from the positive-pair similarity distribution and balances informative negatives with broader coverage. In the two-stage training procedure, Stage 1 learns an encoder from these task-aware pairs, and Stage 2 jointly fine-tunes the encoder and a linear classification head for binary factuality prediction. At inference time, CPIL requires no paraphrase generation, pair construction, or Siamese comparison and uses only one encoder-classifier forward pass. On the 11-task LLM-AggreFact benchmark, CPIL-2% and CPIL-3% achieve macro-F1 scores of 81.04 and 81.60 while using only 2-3% of the FactCG-released labeled pool. Both settings outperform all compared compact detectors and two vanilla LLM references. Ablations and sensitivity analyses support the contributions of task-aware pair construction, dynamic negative mining, and multi-pivot paraphrasing.

Thu 12 NovComputation and LanguageArtificial Intelligence
The gist
Sometimes, computer programs make claims based on documents that are not actually supported by the information in those documents. The authors present a new approach that helps detect whether a claim truly matches the facts in a document but requires very few labeled examples to learn from. Their method creates pairs of claim-document examples by rephrasing claims and comparing ones with opposite labels to teach the model better. It performs well on a standard benchmark while needing much less labeled data and runs efficiently when checking claims.
Open → 2611.16264v1

QueryFlow improves multi-hop evidence search with adaptive queries

QueryFlow: Multi-Hop Graph Retrieval with Adaptive Query Evolution

Abstract: Multi-hop retrieval-augmented generation (RAG) supports Web information seeking and fact verification by recovering complementary evidence across related pages within a limited retrieval budget. Scoring evidence independently against the original question can miss supporting units whose relevance emerges through intermediate facts. Iterative retrieval incorporates these facts into subsequent queries, while graph retrieval explicitly models relations among evidence units. Yet fixed-query graph retrievers update entity representations without adapting the query condition to accumulated relational context. We view multi-hop evidence recovery as the coupled evolution of evidence representations and query conditions, with accumulated context informing both what the retriever represents and how it searches for complementary evidence. QueryFlow realizes this perspective by recurrently integrating query-conditioned summaries of graph context into a shared query state, which controls boundary inputs and relation features for subsequent propagation. Across four representative multi-hop question answering (QA) datasets, QueryFlow achieves macro Recall@10 (R@10), Full Recall@10 (FR@10), and LLM-Acc gains of 4.98, 8.40, and 8.0 percentage points over representative baselines, respectively.

Tue 3 NovInformation RetrievalArtificial Intelligence
The gist
Finding information from multiple related sources can be hard because sometimes clues only make sense when combined. The authors show that treating the search question as something that changes as new facts appear helps find better evidence. Their approach, QueryFlow, updates both the question and the evidence connections step-by-step to find more helpful pages. This method worked better than others on several question answering tests, helping large language models answer more accurately.
Open → 2611.04625v1

Web agents often get wrong content despite 200 OK responses

When 200 OK Is Not an Answer: Silent Fetch Failures in Agentic Web Content Acquisition

Abstract: Web agents and answer engines fetch pages at question time, and liveness checks of cited sources treat an HTTP 200 response as success. We measure how often that assumption fails and what it does to measurements built on fetched pages. We fetched 10,307 URLs in three days with a polite static client and a rendered browser client: pages from 1,507 sites in five strata, random non-existent paths on every site, and the 4,795 URLs cited by expert-verified claims in ExpertQA. We define a taxonomy of what a 200 response contains, label the static 200 responses automatically, and check the labels against 300 items reviewed by an LLM-based assistant; a human check of 60 of them retained every reviewed usability label after discussion. By the automatic labels, 20.9% of the pages that a static client receives with status 200 do not contain the page's main content: app shells, bot challenges, block pages, error and placeholder pages, and partial content. On the held-out test split, the reviewed sample lowers this estimate from 25.5% to 20.0%. 13.1% of the sites whose random-path probe was answered returned 200. The error propagates: for the cited sources, treating 200 as content overstates the share of claim-source pairs that can be checked against live content by 12.2 points and inflates the apparent silent citation-drift rate from 5.8% to 21.9%. Cheap heuristics catch most of it. Text length, probe comparison and redirect-to-root, combined and tuned on held-out sites, recover 86% of unusable reviewed responses at a false-alarm rate near 5%. Content-reading detectors perform comparably: the best, the text classifier, gains five to seven points of recall at more than twice the false-alarm rate. We argue that agents and their evaluations should report fetch outcomes explicitly instead of inferring them from the status line and will release the labelled data and code.

Sat 31 OctInformation RetrievalComputation and LanguageNetworking and Internet Architecture
The gist
Web agents and answer engines check if a webpage is live by looking for a 200 OK HTTP response, assuming it means the page loaded correctly. The authors found that over 20% of these 200 responses actually don’t contain the real page content, but things like error messages or loading shells. This causes problems when relying on these pages for fact-checking or verifying information. They also created simple methods that find most of these failures to improve web content fetching.
Open → 2611.00884v1

Active evidence checking improves fact verification accuracy

Stop or Clarify? Experience-Conditioned Active Fact Verification

Abstract: Fact-verification benchmarks score the final label, while an interactive verifier must first decide whether the available evidence is sufficient. We focus on that intermediate Stop-or-Clarify Decision. Uncertainty alone is insufficient, because the benefit of another evidence-acquisition step is unknown before the step is taken. Sequential Experience-Adaptive Clarification (SEAC) wraps a fixed verifier with a lightweight evidence-acquisition policy. For each claim, SEAC retrieves evidence, runs the verifier, forms a label-free decision state, and combines stop risk with an observational clarification-utility estimate from earlier executed trajectories. Clarification is requested only when this score exceeds the action cost and decision threshold. On SciFact, LIAR, and FEVER, SEAC improves mean accuracy and macro-F1 over the compared fact-verification and active-retrieval baselines. Equal-budget and formula-level diagnostics support selective clarification as a contributor to the gain, while the small SciFact split leaves its paired improvement inconclusive.

Mon 26 OctArtificial Intelligence
The gist
Fact-checking systems often decide whether a claim is true or false based on gathered evidence, but it's unclear when they should stop looking for more information. The authors propose a method called SEAC that helps systems decide if they've found enough evidence or if they should ask for more. This method learns from past experience to balance the benefit of clarifying with the cost of searching further. Tests show SEAC improves accuracy on popular fact-checking tasks by selectively gathering more evidence only when needed.
Open → 2610.34941v1

LLM agents often give inconsistent true and false probabilities for claims

Do "True" and "False" Add Up to One? Complement Consistency in LLM-Agent Opinion Simulations

Abstract: In LLM-agent opinion simulations, each agent reports the probability that a claim is true after seeing its own previous probability and the probabilities reported by its neighbors, and that number becomes its state in the next round. If the number describes the agent's position rather than the sentence it was shown, then rewriting the claim as its complement and replacing every probability p with 1 - p should return the complementary answer, so the two reports should sum to one. We test this with three small open models: Qwen3-4B, Llama-3.2-3B and Gemma-3-4B. On 2,304 agent states taken from a running network simulation, Qwen's paired reports sum to 1.18 on average, and on 48 real fact-checking claims from AVeriTeC they again exceed one. A controlled benchmark of 24 worlds, each with exactly two possible outcomes, shows the same failure in all three models. The discrepancy is two to ten times larger than the change produced by a same-meaning paraphrase, and it persists under repeated sampling and across peer group sizes and report strengths. When the reports are fed back into two populations that start from complementary states, the populations do not remain mirror images: their divergence is positive after five rounds for every model. Averaging each answer with the complement of its paired answer restores consistency by construction and lowers Brier loss averaged over the two wordings, but at the final round it beats the better of the two original wordings for only one of the three models. A schedule-by-targeting experiment on the same model gives a small, stable interaction because such sweeps keep the claim wording fixed. Consistency under rewording therefore needs its own test before simulated probabilities are read as beliefs

Fri 23 OctArtificial IntelligenceComputation and LanguageMultiagent Systems
The gist
When AI agents estimate how likely a statement is true or false, their answers for a claim and its opposite should add up to one. The authors tested three smaller AI language models and found they often do not add up perfectly, sometimes adding to more than one. This inconsistency happens even after multiple rounds of opinion sharing and rewriting the statements. They show a simple fix can reduce this error but does not always improve accuracy. Their work highlights the need to test AI-generated probabilities carefully before treating them as consistent beliefs.
Open → 2610.32874v1

Large language models struggle to spot biases in news articles

BAIT: Can LLMs Find What Biases a Reader? A Benchmark for Cognitive Bias Triggers in News

Abstract: Persuasive writing exploits cognitive biases, recruiting a reader's mental shortcuts to shape belief, often without the reader noticing. We call the linguistic features exploiting such mental shortcuts cognitive bias triggers. As large language models (LLMs) increasingly mediate how people encounter news, summarizing it, ranking it, and answering questions over it, a natural question is whether they can recognize the features of a text that bias a human reader. We introduce BAIT (Bias Annotation In Text), a resource for detecting cognitive bias triggers in news. BAIT pairs a schema of nine triggers, each specified in an annotator codebook, with 50 English news articles annotated by trained annotators at the article and span levels and adjudicated into a gold standard. On this data we define two tasks, detecting which triggers a text contains and locating the spans that carry them. We then benchmark six frontier LLMs zero-shot and with three or ten in-context examples. The models approach human agreement on which triggers an article contains, but rarely recover its full set of triggers, and they reach less than half of human agreement in locating the passages that carry them. Moreover, they over-predict triggers that come with easy surface cues, and most of the passages they mark for these triggers are not triggers, yet carry the trigger's surface cue far more often than random text. These findings suggest that the evaluated LLMs respond to surface cues rather than to the mechanisms that make a passage manipulative. We release the schema, the codebook, standoff annotations, and the evaluation pipeline.

Tue 20 OctComputation and LanguageArtificial Intelligence
The gist
Persuasive news writing often uses tricks that tap into mental shortcuts called cognitive biases to influence readers without them noticing. The authors created a dataset called BAIT containing news articles labeled with nine specific types of these bias triggers so models can learn to find them. They tested six advanced language models, which could mostly tell if an article had some bias triggers but often missed many and struggled to pinpoint the exact biased parts. The models also tended to flag common surface words rather than truly understanding the manipulative intent behind the text. This suggests these models aren't yet good at recognizing the real causes of bias in writing.
Open → 2610.28092v1

Large language model agents improved with professional fact-checking skills

Standing on the Shoulders of Giants: Enhancing LLM Agents for Automated Fact-Checking with Human Expert Skills

Abstract: Automated fact-checking (AFC) systems powered by large language model (LLM) agents still face challenges in evidence retrieval and claim verification. Existing approaches improve their capabilities through computationally costly post-training or human-in-the-loop feedback that is difficult to scale. To transfer professional fact-checking expertise to LLM agents efficiently, we distill publicly available guidelines published by Agence France-Presse (AFP) Fact Check into a reusable, modular Professional Fact-Checking Skill (PFC-Skill). We then present SkillCheck, the first AFC system that gives LLM agents dynamic access to PFC-Skill to guide evidence retrieval and claim verification at inference time. Extensive experiments demonstrate that SkillCheck improves upon skill-free DEFAME baselines across multiple LLM backbones, with gains of 2.2--4.8 pp. in Accuracy and 1.5--6.3 pp. in Macro-F1 on the AVeriTeC benchmark. Notably, SkillCheck with GPT-4o-Mini exceeds DEFAME with GPT-4o by 3.4 pp. in Accuracy while cutting inference cost by roughly 61%. The performance gains also extend to open-web retrieval. Ablations show that incorporating the skill during evidence retrieval provides most of the gains, while incorporation into both stages yields the best overall performance. Case studies show that PFC-Skill enables more targeted search and closer scrutiny of evidence. To conclude, our work shows that LLM agents can "stand on the shoulders of giants" by drawing on the best practices of professional fact-checkers, substantially enhancing AFC performance without extra training or continuous human supervision.

Mon 19 OctArtificial IntelligenceComputation and Language
The gist
Automated fact-checking tools using large language models still struggle to find and verify evidence well. The authors created a special skill module based on real expert fact-checking guidelines to help these tools work better. By letting the language models access this skill during their checking process, the system, called SkillCheck, improves accuracy while reducing computer costs. This approach helps language models search smarter and examine evidence more carefully without needing extra training or constant human help.
Open → 2610.25777v1

Large language models accurately detect covid19 fake news on twitter

A Large Language Model LLM Approach to COVID-19 Fake News Detection

Abstract: Text classification is a widely studied problem in information retrieval and data mining, with applications ranging from healthcare diagnosis to the detection of COVID-19-related fake news. Social media platforms generate vast volumes of user-generated text, and the rapid spread of misinformation during the pandemic has made automated fake news detection an important task. In this study, we apply pretrained transformer-based transfer learning (TL) models and large language models (LLMs), namely BERT-base, ALBERT-base, XLM-RoBERTa-base, DistilBERT, GPT-2, XLNet, and T5, to the COVID-19 English Tweets dataset, which contains 6,420 tweets labeled as real or fake. To further improve performance, we apply six late-fusion optimization techniques: weighted fusion, particle swarm optimization (PSO) fusion, the Nelder-Mead algorithm, Powell's method, the Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm, and trust-region constrained optimization. Models are evaluated using accuracy, precision, recall, and F1-score, and confusion-matrix heatmaps are used to visualize the predictions of each model. GPT-2 achieved the best performance, with an accuracy of 0.9631, precision of 0.9612, recall of 0.9610, and F1-score of 0.9600, while Nelder-Mead fusion achieved the best result among the fusion methods with an accuracy of 0.9610. These results highlight the value of LLMs for COVID-19 fake news detection. We conclude by outlining future directions, including multiclass classification, multilingual datasets, and real-time social media streams.

Fri 16 OctComputers and Society
The gist
Misinformation about COVID-19 spreads quickly on social media, so spotting fake news automatically is important. The authors tested several advanced language models on a dataset of English tweets labeled as real or fake. They found that GPT-2 was the most accurate for identifying fake COVID-19 news tweets. They also improved results by combining model predictions using several fusion techniques. This shows that large language models can help filter misinformation during a pandemic.
Open → 2610.23800v1

Models identify American satire more accurately than Australian satire

Unreliable Narrator Framing as a Cost-effective Mechanism for Satire Explanation of Australian and American Satirical News

Abstract: Satirical news is a unique form of misinformation: it is intended to be humorous by the original publication source but can be misinterpreted. Explanation of satirical news articles is challenging because recent and local context needs to be incorporated. This is a pilot study that explores explanation generation for satirical news articles from Australia and compares them alongside those from USA. The paper demonstrates that framing explanations on unreliable narrator theory on open-weights models can be a feasible, cost-effective alternative to frontier models. We find that models identify the satirical tropes more accurately on the American source than the Australian one. This work results in a novel dataset and an evaluation benchmark for satirical explanation.

Fri 9 OctComputation and Language
The gist
Satirical news is meant to be funny but can sometimes confuse readers who don't get the joke. The authors explored a way to explain why satire works using a theory about unreliable narrators, which are storytellers who can't be fully trusted. They tested this idea on satirical news from Australia and the USA using affordable language models. Their study found that the models were better at recognizing satire in American articles than Australian ones. They also created a new collection of data and a way to measure how well satire explanations work.
Open → 2610.12863v1

Large language models improved for numerical claim checks

Towards Robust Numerical Claim Verification

Abstract: Large language models (LLMs) are widely used for claim verification, yet remain brittle for numerical reasoning: even small changes in value can sharply degrade accuracy. We show that this brittleness persists in frontier LLMs, but can be mitigated through adversarial fine-tuning on numerically perturbed examples. Using parameter-efficient fine-tuning, small Qwen3 models (0.6B$\unicode{x2013}$8B) reach 98.7% accuracy on label-flipping perturbations, outperforming larger zero-shot models and frontier systems (GPT-5.4 Pro (74.0%) and Gemini 2.5 Flash (73.9%)). The gains generalise to unseen perturbation types, indicating robust numerical decision boundaries rather than memorised edits. Robustness also transfers without target-domain data, significantly improving cross-lingual performance in Spanish. We further show that the same fine-tuning recipe confers robustness to evidence-side perturbations, using the VitaminC dataset.

Wed 30 SeptComputation and Language
The gist
Verifying claims that include numbers is tricky for large language models (LLMs) because even small number changes can cause mistakes. The authors found that this problem still happens in the newest LLMs but can be fixed by fine-tuning them with examples where numbers are deliberately changed. This training method helps smaller models outperform bigger, popular ones and even works well in different languages without extra data. Their approach also improves handling cases where the evidence itself has tricky changes.
Open → 2610.00689v1

Multimodal misinformation detection improves with tool guided learning

MM-VeriAgent: Learning to Use Extensive Tools to Verify Multimodal Misinformation with Reinforcement Learning

Abstract: Real-world multimodal misinformation often involves mixed forgery sources, requiring sample-specific detection strategies. Existing tool-augmented methods rely on predefined workflows or inference-time planning, limiting adaptability or increasing inference cost. To address this issue, we introduce \textbf{MM-VeriAgent}, which learns to verify mixed-source multimodal misinformation with tools. We first build \textbf{MM-VeriTools}, a specialized toolkit for misinformation detection agents. By benchmarking various candidate models and methods on the sub-tasks required by mixed-source detection, we select the strongest for textual, visual, and cross-modal forgery analysis and encapsulate them as callable tools with a unified interface. On top of this toolkit, we train the LVLM agent with reinforcement learning to teach it how to use these tools to better solve mixed-source detection. Since many of the tools are specialized models whose online execution at every rollout severely limits RL efficiency, we further introduce \textbf{Tool-Execution Cache}, which pre-executes candidate tool calls and reuses their cached outputs during training. This preserves multi-step rollouts while reducing online tool execution, largely improving the training efficiency.Experiments on MMFakeBench demonstrate substantial accuracy gains over the base model without explicit tool search at inference time. Ablation and efficiency analyses further validate the learned tool-use policy and show that Tool-Execution Cache reduces online tool executions during training.

Fri 25 SeptComputer Vision and Pattern Recognition
The gist
Detecting false information that mixes text and images is tricky because it can come from multiple kinds of fakes. The authors created a toolkit of specialized models for checking text, images, and their combinations. They then trained an AI agent to learn how to use these tools together efficiently, improving its ability to spot mixed-source misinformation. A smart caching system also sped up training by reusing previous results instead of running all checks repeatedly.
Open → 2609.30698v1

Web tool uses argument schemes and ai to analyse deception in texts

DeceptionAnalyser: A Web-Based AI Tool for Performing Structured Deception Analysis with Argumentation Schemes and LLMs

Abstract: Deception plays a central role in Intelligence operations, yet it remains difficult to analyse systematically without expert knowledge of reasoning patterns and cognitive manipulation. In computational argumentation, for instance, no scheme-level ground-truth corpora currently exist to support statistical validation. In this paper, we address this by introducing a set of ten argument schemes designed to model distinct forms of deception, each accompanied by structured premises and critical questions. In doing so, we introduce the first dedicated library of argumentation schemes specifically designed for deception analysis, providing a structured foundation for systematically modelling and analysing deception in narrative text. We then present \textit{DeceptionAnalyser}, a browser-based tool that implements these schemes through a two-stage methodology combining LLM-based premise extraction with critical-question-driven evaluation. Our aim is to provide a conceptual and methodological foundation for analysing deceptive reasoning in narrative text. This is precisely what we address in this paper by demonstrating how structured argumentation theory and AI-assisted analysis can support transparent, explainable assessments of potential deception. Because the schemes are designed to flag claims for scrutiny rather than to output a deception verdict, we do not benchmark classification accuracy; instead, we assess the \emph{reliability} of the methodology by measuring the consistency of the tool's premise and conclusion assessments across ten contemporary large language models and repeated runs. We find that scheme detection is highly stable for clear-cut deception and degrades gracefully, in interpretable ways, on more ambiguous intelligence-style narratives.

Mon 21 SeptHuman-Computer InteractionArtificial Intelligence
The gist
Detecting lies and trickery in stories or statements is very hard, especially without experts who know how deception works. The authors created ten patterns that model different kinds of deception, making it easier to spot when claims need closer checking. They built a web tool called DeceptionAnalyser that uses artificial intelligence to find important parts of a statement and then asks smart, critical questions based on those patterns to evaluate if something seems suspicious. Instead of saying outright if something is a lie, the tool highlights what needs more attention, helping people understand why a claim might be deceptive. They tested this method across many AI models and found it works well for clear cases and degrades gracefully when things are less obvious.
Open → 2609.24369v1

Formal methods clarify automated fact checking and accountability

SoK: Formal Methods for Fact-Checking and Information Integrity

Abstract: An automated fact-checking system returns a label: the claim is true, or it is false. In many such systems the verdict remains the primary output. What is generally missing is a record of which document settled the question, of what would have had to be different for the verdict to change, or of whether the same claim, reworded, would have been judged the same way. We call the missing piece a warrant: a separate statement of what was guaranteed and on what grounds. Formal methods produce evidence of this kind, and regulation is beginning to ask for it, since the Digital Services Act and the AI Act both call for auditable evidence about how systems behave. Surveys of automated fact-checking are usually organised by pipeline stage, and treat logic as one technique among many. We organise the field by what is being formalised instead, which gives five levels: the claim, the reasoning, the system doing the checking, the ecosystem the claim spreads through, and the regulatory obligation. Sorting 121 works into those levels, two patterns stand out. Most of the relevant formal machinery already exists, but it was built for other domains and has rarely been applied here, and the gap is widest for verifying the checking system itself. Several stages of the routine professional fact-checkers follow also have no stated correctness criterion, and two of them, writing a claim in checkable form and correcting a verdict already published, are not formally specified in any work we coded. We close with open problems, each with a suggested first step.

Sat 19 SeptComputation and LanguageComputers and SocietyLogic in Computer Science
The gist
Fact-checking systems often give simple true or false answers but don’t explain exactly how they reached those conclusions or what evidence supports them. The authors look at ways formal methods—strict mathematical techniques—can provide clear evidence for fact-checkers’ decisions. They organize 121 studies by what parts of fact-checking are formalized, finding some important gaps such as defining correctness for certain steps. Their work highlights what research exists, what is missing, and suggests next steps to make fact-checking more transparent and reliable.
Open → 2609.23239v1

Llms used for fact checking rely more on memory than evidence

Evaluating and Improving Evidence-Grounded Fact-Checking in LLMs via Multi-Round Evidence Ablation

Abstract: Automatic fact-checking systems assess the veracity of claims given evidence from relevant documents. Large Language Models (LLMs) have demonstrated strong performance in fact-checking due to their general reasoning capabilities. However, it remains unclear whether they faithfully make use of the evidence provided to reach veracity judgments or rely on parametric knowledge. To investigate this, we introduce Fact-Ablated Evaluation (FAE), a new evaluation framework that iteratively ablates the cited evidence to assess whether LLMs revise their predictions accordingly. Our empirical results show that current off-the-shelf LLMs as fact-checking systems rely more on their parametric knowledge than on the evidence provided. To bridge this gap between prediction accuracy and evidence grounding, we propose REAL (Rigorous Evidence Ablation Learning), a training framework that promotes evidence-dependent verification through counterfactual evidence supervision for the LLM-as-verifier models. Experiments on four fact-checking datasets across different domains demonstrate that models trained with REAL obtain superior evidence-dependent capabilities compared to standard fine-tuned models. Our findings highlight that strong fact-checking performance can still coexist with weak evidence dependency, while REAL encourages veracity predictions to remain more closely tied to the availability of supporting evidence.

Tue 8 SeptComputation and LanguageArtificial IntelligenceInformation Retrieval
The gist
Fact-checking systems aim to verify claims using evidence, but it is unclear if large language models (LLMs) actually use that evidence or just rely on what they already know. The authors created a test called Fact-Ablated Evaluation that gradually removes pieces of evidence to see if LLMs change their answers. They found that many LLMs tend to ignore the evidence and depend mostly on their stored knowledge. To fix this, they introduced a training method called REAL to teach LLMs to pay closer attention to the evidence when verifying facts. Their experiments show this method helps models rely more on provided evidence while still making accurate judgments.
Open → 2609.08943v1