A Graph-Enhanced Defense Framework for Explainable Fake News Detection with LLM
Authors: Bo Wang, Jing Ma, Hongzhan Lin, Zhiwei Yang, Ruichao Yang, Yuan Tian, Yi Chang
Organizations: School of Artificial Intelligence, Jilin University, China · Hong Kong Baptist University, China · Guangdong Institute of Smart Education, Jinan University, China · University of Science and Technology Beijing, China · Engineering Research Center of Knowledge-Driven Human-Machine Intelligence, China · International Center of Future Science, Jilin University, China
Explainable fake news detection aims to assess the veracity of news claims while providing human-friendly explanations. Existing methods incorporating investigative journalism are often inefficient and struggle with breaking news. Recent advances in large language models (LLMs) enable leveraging externally retrieved reports as evidence for detection and explanation generation, but unverified reports may introduce inaccuracies. Moreover, effective explainable fake news detection should provide a comprehensible explanation for all aspects of a claim to assist the public in verifying its accuracy. To address these challenges, we propose a graph-enhanced defense framework (G-Defense) that provides fine-grained explanations based solely on unverified reports. Specifically, we construct a claim-centered graph by decomposing the news claim into several sub-claims and modeling their dependency relationships. For each sub-claim, we use the retrieval-augmented generation (RAG) technique to retrieve salient evidence and generate competing explanations. We then introduce a defense-like inference module based on the graph to assess the overall veracity. Finally, we prompt an LLM to generate an intuitive explanation graph. Experimental results demonstrate that G-Defense achieves state-of-the-art performance in both veracity detection and the quality of its explanations.
Figures & tables
Figure 1. A false claim from Sina Weibo. Left: The comparison of informativeness and soundness between two competing parties serves as an indicator of veracity. Right: The different modeling of decomposed sub-claims.
Figure 2. An architecture overview of the proposed G raph-enhanced Defense framework for explainable fake news detection ( G-Defense ). Given a claim c , the graph construction module first constructs a claim-centered graph through claims decomposition and edge generation, where the claim is decomposed into n sub-claims. Each sub-claim ci is then individually processed to retrieve evidence and generate competing explanations. Next, the claim-centered graph enriched with explanations is used for defense-like inference to derive the veracity prediction of the claim. Finally, an explanation graph and a summarized explanation T can be derived from the last module.
Figure 3. An illustrative example of the graph construction process.
Figure 4. An illustrative example of the competing explanation generation process for a sub-claim.
Dataset
RAWFC
LIAR-RAW
Claim
# Claims
2,012
12,590
# pants-fire
-
1,013
# false
646
2,466
# barely-true
-
2,057
# half-true*
671
2,594
Table 1. Summary statistics of datasets. # half-true* is also denoted as # half in RAWFC.
RAWFC
LIAR-RAW
P
R
macF1
P
R
macF1
Traditional approach
dEFEND ( Shu et al., 2019 ) †
44.93
43.26
44.07
23.09
18.56
17.51
SBERT-FC ( Kotonya and Toni, 2020a ) †
51.06
45.92
45.51
24.09
22.07
22.19
GenFE ( Atanasova et al., 2020 ) †
44.29
44.74
44.43
28.01
26.16
26.49
GenFE-MT ( Atanasova et al., 2020 ) †
45.64
45.27
45.08
18.55
19.90
15.15
Table 2. Veracity prediction results on RAWFC and LIAR-RAW. † Resulting numbers are reported by Yang et al. (2022b) , the results of FactLLaMA and L-Defense are taken from the original paper, and the results of MUSER are reproduced by us. Bold numbers indicate the best performance, and underlined numbers indicate the second-best performance. Results for G-Defense and its variants are averaged over three random runs.
Figure 5. Distribution of claims by the number of decomposed sub-claims in the test sets of RAWFC and LIAR-RAW.
RAWFC
LIAR-RAW
Method
P
R
macF1
P
R
macF1
G-Defense
65.89
65.54
65.65
32.94
32.09
32.09
w/o sub-claims
56.29
56.02
56.10
30.94
28.13
28.52
w/o edges
64.12
63.49
63.40
31.39
30.52
30.60
w/o evidence
56.68
55.98
55.76
30.12
29.44
29.30
w/o competing
57.24
55.96
55.94
30.42
29.53
29.46
Table 3. Ablation study of veracity prediction on two datasets.
Figure 6. MacF1 scores of G-Defense and its w/o-sub-claims variant grouped by the number of sub-claims on the RAWFC (left) and LIAR-RAW (right) test sets. The percentages under each group indicate the proportion of samples within that group.
RAWFC
LIAR-RAW
D ↓
M ↓
I ↑
S ↑
R ↑
D ↓
M ↓
I ↑
S ↑
R ↑
Oracle
-
1.52
4.46
4.73
4.72
-
1.85
4.44
4.60
4.69
CofCED ( Yang et al., 2022b )
1.53
2.74
2.89
1.93
2.46
1.33
3.64
1.75
1.76
1.59
GPT-3.5 claim
1.70
1.97
4.00
4.44
4.68
1.39
2.27
3.93
4.29
4.50
GPT-3.5 full
1.81
2.07
4.44
4.62
4.69
1.39
2.29
3.71
4.04
3.99
L-Defense LLaMA2 ( Wang et al., 2024 )
1.30
1.95
4.44
4.67
4.62
1.36
2.20
4.39
4.64
4.63
Table 4. Evaluation results of explanation quality using a 5-Point Likert scale rating by ChatGPT on RAWFC and LIAR-RAW. For metrics D and M, a lower score ( ↓ ) indicates better performance, while a higher score ( ↑ ) indicates better performance for the remaining metrics. The bold numbers denote the best results in addition to Oracle while the underlined ones are better than Oracle.
Figure 7. Explanation evaluation results using a 5-Point Likert scale rating by five human annotators on 30 randomly sampled claims from the test set of RAWFC. Scores from five annotators were averaged.
Figure 8. Confusion matrices of judgment results from five annotators on 30 randomly sampled claims. The three results are collected under three different settings: annotators are provided with only the claim, with the claim and explanations generated by L-Defense, or with the claim and explanations generated by G-Defense. The results from five annotators were averaged and rounded for visualization.
claim 1 ( True ): A newly discovered species of amphibian was named “Dermophis donaldtrumpi” as an unflattering reference to the U.S. President.
[1pt/1pt]
c1 ( True ): A newly discovered species of amphibian was named “Dermophis donaldtrumpi”. c1+ : The newly discovered species of amphibian was named “Dermophis donaldtrumpi” in recognition of the US presidentś climate change denial. c1− : The claim is false because there is no newly discovered species of amphibian named “Dermophis donaldtrumpi” c2 ( True ): The name “Dermophis donaldtrumpi” was chosen as an unflattering reference to the U.S. President. c2+ : The name “Dermophis donaldtrumpi” was chosen as an unflattering reference to the U.S. President because EnviroBuild, who paid for the naming rights, stated that they chose the name in recognition of the presidentś position on climate change. c2− : The name “Dermophis donaldtrumpi” was chosen as a reference to the U.S. Presidentś climate change policy.
Summary : The claim is reasoned as ‘ true ’ based on the supporting evidence provided by the true-oriented explanations for both sub-claims, which establish a direct connection between the naming of the species and the U.S. President’s stance on climate change.
claim 2 ( Half ): The U.S. Congress passed the 22nd Amendment in order to “make sure President Franklin D. Roosevelt did not get re-elected”.
[1pt/1pt]
c1 ( True ): The U.S. Congress passed the 22nd Amendment. c1+ : The U.S. Congress passed the 22nd Amendment in 1947, which was later ratified by the states in 1951. c1− : The 22nd Amendment was approved by Congress in 1947, with an exception that would exclude a president in office from term limit during the ratification process. c2 ( False ): The purpose of the 22nd Amendment was to prevent President Franklin D. Roosevelt from being re-elected. c2+ : The 22nd Amendment was not intended to prevent President Franklin D. Roosevelt from being re-elected, but rather to establish a term limit for all future presidents. c2− : The 22nd Amendment was not intended to prevent President Franklin D. Roosevelt from being re-elected, but rather to establish a term limit for all future presidents. FDR died in 1945. The amendment was passed in 1947 and ratified in 1951 — after his death.
Summary : The claim is reasoned as ‘ half ’ because while it is true that the U.S. Congress passed the 22nd Amendment, the assertion that it was solely to prevent President Franklin D. Roosevelt from being re-elected is false. The actual purpose was to set a term limit for all future presidents.
claim 3 ( False ): By ignoring a ban on performing COVID-19 autopsies, researchers in Italy discovered that the disease is caused by a bacteria — not a virus — and can simply be treated with aspirin.
Table 5. Case study. The claims are selected from the test set of RAWFC. We provide both competing explanations for analysis. In practice, only the explanation corresponding to the predicted label would be presented to users, and the other one (shown in gray) is included here for analysis. The gold label of the claim and the predicted labels of both the claim and its sub-claims are highlighted in color.
Table 6. Token usage (in millions) and estimated cost of G-Defense on LIAR-RAW and RAWFC across all LLM-driven stages.
Propagation structures provide crucial evidence for fake news detection, yet existing approaches primarily rely on supervised GNN-based models, which require substantial labeled data and exhibit limited generalization. Although large language models (LLMs) exhibit strong reasoning capabilities, directly feeding them raw propagation graphs creates a significant modality mismatch and severe information overload, making structure-aware reasoning unreliable in zero-shot and few-shot settings. To bridge this gap, we propose MAGER, a multi-agent genetic evolution framework that automatically discovers meta-paths optimized for LLM reasoning. By compressing complex propagation graphs into informative subgraphs, the evolved meta-paths alleviate both information overload and modality mismatch, enabling frozen LLMs to perform structure-aware veracity reasoning. We further introduce a graph in-context learning strategy that retrieves semantically and structurally similar demonstrations to strengthen classification and reasoning. Extensive experiments show that MAGER substantially improves frozen LLMs as standalone fake news detectors in data-efficient settings. Our code is available at https://github.com/SenticNet/MAGER.
Ziyi Zhou, Xiaoming Zhang, Hui Pang +5
Beihang University Beijing, China · Chinese Academy of Sciences Beijing, China · Beijing University of Posts and Telecommunications Beijing, China +2
TRUST Agents is a collaborative multi-agent framework for explainable fact verification and fake news detection. Rather than treating verification as a simple true-or-false classification task, the system identifies verifiable claims, retrieves relevant evidence, compares claims against that evidence, reasons under uncertainty, and generates explanations that humans can inspect. The baseline pipeline consists of four specialized agents. A claim extractor uses named entity recognition, dependency parsing, and LLM-based extraction to identify factual claims. A retrieval agent performs hybrid sparse and dense search using BM25 and FAISS. A verifier agent compares claims with retrieved evidence and produces verdicts with calibrated confidence. An explainer agent then generates a human-readable report with explicit evidence citations. To handle complex claims more effectively, we introduce a research-oriented extension with three additional components: a decomposer agent inspired by LoCal-style claim decomposition, a Delphi-inspired multi-agent jury with specialized verifier personas, and a logic aggregator that combines atomic verdicts using conjunction, disjunction, negation, and implication. We evaluate both pipelines on the LIAR benchmark against fine-tuned BERT, fine-tuned RoBERTa, and a zero-shot LLM baseline. Although supervised encoders remain stronger on raw metrics, TRUST Agents improves interpretability, evidence transparency, and reasoning over compound claims. Results also show that retrieval quality and uncertainty calibration remain the main bottlenecks in trustworthy automated fact verification.
This article examines the application of Explainable Artificial Intelligence (XAI) in NLP based fake news detection and compares selected interpretability methods. The work outlines key aspects of disinformation, neural network architectures, and XAI techniques, with a focus on SHAP, LIME, and Integrated Gradients. In the experimental study, classification models were implemented and interpreted using these methods. The results show that XAI enhances model transparency and interpretability while maintaining high detection accuracy. Each method provides distinct explanatory value: SHAP offers detailed local attributions, LIME provides simple and intuitive explanations, and Integrated Gradients performs efficiently with convolutional models. The study also highlights limitations such as computational cost and sensitivity to parameterization. Overall, the findings demonstrate that integrating XAI with NLP is an effective approach to improving the reliability and trustworthiness of fake news detection systems.
Krzysztof Siwek, Daniel Stankowski, Maciej Stodolski
Warsaw University of Technology, Warsaw, Poland · Student of Warsaw University of Technology, Warsaw, Poland