RAEGNet: Relation-Aware Evidence Graph Network for Harm-Aware Multimodal Fake News Detection

Wenbin Shen; Guoxuan Qin; Guangxu Yao; Baodong Wang; Yuanbo Rui; Zhongjie Ba; Zhichao Lian

cs.CLSep 29, 2026

RAEGNet: Relation-Aware Evidence Graph Network for Harm-Aware Multimodal Fake News Detection

Authors: Wenbin Shen, Guoxuan Qin, Guangxu Yao, Baodong Wang, Yuanbo Rui, Zhongjie Ba, Zhichao Lian

Organizations: Nanjing University of Science and Technology · Zhejiang University · University of Chinese Academy of Sciences

Abstract

Existing multimodal fake news detection methods often introduce external information to assist detection. However, most of them rely on entity-level retrieval and are therefore prone to introducing event-irrelevant noise. Meanwhile, existing methods mainly focus on improving overall performance and do not account for differences in the degree of harm posed by different instances of fake news. To address these limitations, we design an Event-Level Evidence Retrieval Framework (ELERF) and propose a Relation-Aware Evidence Graph Network (RAEGNet). ELERF retrieves external evidence based on the complete event semantics of a news item. RAEGNet constructs a directed graph that incorporates news-evidence stance relations and evidence-evidence interaction relations, and introduces a conditional-harm branch to jointly model authenticity and potential harm. Experimental results demonstrate that RAEGNet outperforms multiple baseline methods across all evaluated metrics on Weibo-21, Fakeddit, and our self-constructed SSS dataset.

Figures & tables

Figure 1: Comparison between entity-level retrieval and our event-level retrieval.

Figure 2: The overall architecture of RAEGNet.

Components	Weibo-21	Fakeddit	SSS
News	4,493	5,919	7,997
Evidence	8,341	11,375	15,157

Table 1: Numbers of news and automatically retained evidence on the three datasets.

Category	Method	Weibo-21			Fakeddit			SSS
		Accuracy	F1-score		Accuracy	F1-score		Accuracy	F1-score
		Accuracy	Fake	Real	Accuracy	Fake	Real	Accuracy	Fake	Real
w/o EK	CAFE	0.815	0.810	0.820	0.786	0.816	0.745	0.726	0.753	0.693
	MRML	0.903	0.908	0.898	0.860	0.876	0.839	0.758	0.779	0.732
	Event-Radar	0.881	0.884	0.877	0.840	0.859	0.815	0.754	0.767	0.739
	MSACA	0.894	0.900	0.886	0.846	0.864	0.823	0.762	0.778	0.744

Table 2: Overall performance comparison of RAEGNet and baselines on the three datasets. “w/o EK” denotes methods that do not use external knowledge, “w/ EK” denotes methods that incorporate external knowledge, and “w/ LLM” denotes LLM-based methods. RAEGNet-E, RAEGNet-B, and RAEGNet-A use entity-level evidence, pre-news event-level evidence, and full event-level evidence, respectively.

Ablation Settings	Weibo-21			Fakeddit			SSS
	Accuracy	HHF		Accuracy	HHF		Accuracy	HHF
	Accuracy	F1	Rec.	Accuracy	F1	Rec.	Accuracy	F1	Rec.
w/o Evidence Graph	0.913	0.956	0.961	0.909	0.939	0.952	0.822	0.892	0.956
– w/o Edge Weights	0.920	0.959	0.956	0.915	0.936	0.935	0.837	0.886	0.923
– w/o Edge Types	0.914	0.960	0.967	0.911	0.930	0.935	0.826	0.877	0.928
w/o Harm-aware Branch	0.915	0.946	0.919	0.912	0.925	0.907	0.839	0.878	0.905

Table 3: Ablation study on the Weibo-21, Fakeddit, and SSS datasets. Indented rows prefixed by “–” indicate that only the specified subcomponent of the corresponding higher-level module is ablated.

Figure 3: Detection performance of RAEGNet at different harm levels.

Dataset	Retrieval	Acc.	Fake F1	Real F1
Weibo-21	Entity-level	0.911	0.911	0.911
Weibo-21	Event-level	0.931	0.934	0.929
Fakeddit	Entity-level	0.898	0.908	0.886
Fakeddit	Event-level	0.919	0.926	0.910
SSS	Entity-level	0.808	0.817	0.798
SSS	Event-level	0.843	0.860	0.822

Table 4: Comparison between entity-level and event-level evidence retrieval under the same RAEGNet architecture.

Figure 4: Sensitivity analysis of

\lambda_{\mathrm{harm}}

and

\lambda_{\mathrm{rank}}

Appendix figures & tables27 assets

Supplementary material from the paper’s appendix.

Appendix

Role	Content
System	You are an expert in fake news research data screening, dedicated to filtering high-quality multimodal news data.
	Each input contains a news text and its associated news image. Evaluate the two modalities jointly.

	I. Judgment Criteria:
	1. Applicable conditions (must be fully met to return “Usable”):
	(1) The text must possess a clear news format and complete structure. The selected content cannot be fragmented isolated words, but must have a clear description of the news event, and contain the basic elements constituting the news.

Appendix

Table 5: The structured system prompt and examples for LLM-based semantic filtering.

Dataset	Language	Original Quantity	Filtered Quantity
Fakeddit ( Nakamura et al., 2020 )	English	1,063,106	2,922
FineFake ( Zhou et al., 2026 )	English	16,909	2,461
MR2 ( Hu et al., 2023 )	Chinese & English	14,700	442
Weibo-17 ( Jin et al., 2017 )	Chinese	9,528	942
Weibo-21 ( Nan et al., 2021 )	Chinese	9,128	850
CFND ( Zhang et al., 2024b )	Chinese	26,665	380

Appendix

Table 6: Specific distribution and volume changes of each data source before and after screening.

Dataset	Samples	Candidate Pairs	Verified Event Groups	Grouped Samples
Weibo-21	4,493	286	74	198
Fakeddit	5,919	413	96	271
SSS	7,997	672	143	406

Appendix

Table 7: Statistics of similarity-assisted event grouping used for fold construction. Candidate pairs are retrieved automatically, whereas event identity is determined by human reviewers.

Harm label	Level	Typical content
0.1	Very low harm	Entertainment gossip, minor misleading content, and everyday misinformation with a limited scope of influence.
0.3	Low harm	Content that may cause localized misunderstanding, but the impact is limited and the harm is controllable.
0.5	Moderate harm	Content involving public events, social controversies, or collective perceptions that may provoke disputes or cause a certain degree of adverse impact.
0.7	High harm	Content involving public safety, health, politics, disasters, or intergroup conflict that may produce a clear social impact.
0.9	Very high harm	Content that may induce panic, real-world harm, risks to public order, or large-scale collective impact.

Appendix

Table 8: Annotation criteria for harm levels.

Evaluation	Weibo-21	Fakeddit	SSS
Human annotator agreement (Krippendorff’s $\alpha$ )	0.9598	0.9303	0.9413
LLM–human agreement (mean agreement)	92.02%	89.37%	90.13%
Adjudication Trigger Rate	0.78% (35/4,493)	1.32% (78/5,919)	1.25% (100/7,997)

Appendix

Table 9: Agreement and adjudication rates for harm labels.

Figure 6: Distribution of harm coefficients in the SSS dataset.

Primary Category	Secondary Category	Retrieval Domains
Encyclopedias and Knowledge Bases		wikipedia, britannica, baidu
Communities and Content Platforms		zhihu, douban, reddit, quora, youtube, weibo, sina, tencent, 163, sohu
News Media	News Agencies	reuters, apnews, xinhuanet, chinanews, tass
	Transnational Media	bbc, cnn, aljazeera, dw, euronews, france24, rt, sputniknews, chinadaily, sky, ifeng
	National Media	cctv, people, thepaper, huanqiu, guancha, gmw, nytimes, washingtonpost, theguardian, foxnews, cbsnews, nbcnews, npr, pbs, time, newsweek, independent, dailymail, usatoday, telegraph, express, mirror, huffpost, cbc, theglobeandmail, ndtv, thehindu, hindustantimes, indianexpress, indiatoday, straitstimes, japantimes, haaretz, timesofisrael, jpost, elpais, nzherald, buzzfeednews, vice, vox, slate, theatlantic, buzzfeed
	Local and Regional Media	bjnews, oeeee, scmp, chicagotribune, boston, bostonglobe, miamiherald, syracuse, denverpost, seattletimes, dallasnews, houstonchronicle, tampabay, kansascity, cleveland, oregonlive, sltrib, latimes, bangordailynews, nola, jacksonville, mercurynews, thestar, torontosun, vancouversun, smh, abc7, abc7ny, ktla, cbslocal, abc13, 9news, whyy

Appendix

Table 10: Retrieval domains organized by primary and secondary categories.

Dataset	Temporal Subset	Candidates	After Explicit Filtering	After LLM Screening
Weibo-21	Pre-news	8,193	6,448	4,094
Weibo-21	Post-news	9,894	7,322	4,247
Fakeddit	Pre-news	11,727	8,456	6,757
Fakeddit	Post-news	13,791	7,333	4,618
SSS	Pre-news	18,526	12,324	8,606
SSS	Post-news	21,718	9,167	6,551

Appendix

Table 12: Temporal composition and leakage-control statistics of the retrieved evidence. The final automatic evidence set corresponds to the evidence retained after LLM screening.

System Prompt

You are a rigorous semantic stance annotator for news-evidence pairs. You need to judge the semantic relation between the evidence and the core event, subject, and claim expressed by the news. Please assign a continuous stance score in the interval

[0,1]

to each evidence item. Interpret the score using the following ranges: •

[0.0,0.2]

Strong refutation: the evidence directly conflicts with the core claim of the news, or explicitly provides facts opposite to the core content of the news. •

(0.2,0.4]

Weak refutation: the evidence is partially inconsistent with the news, or weakly negates or weakens a key detail of the news. •

(0.4,0.6]

Neutral: the evidence is only background-related or topic-related, provides insufficient information, cannot determine support or refutation, or is basically irrelevant to the news. •

(0.6,0.8]

Weak support: the evidence is partially consistent with the news, or provides weak support for a key detail of the news. •

(0.8,1.0]

Strong support: the evidence explicitly supports the core claim of the news, or provides facts highly consistent with the core content of the news. Important constraints: • Support or refutation only indicates the semantic relation between the evidence text and the news text, and is independent of the actual truthfulness of the news. • Do not complete missing facts using commonsense or external knowledge. Make the judgment only based on the input text. Output requirement: Return only one decimal number in

[0,1]

, without explanations or additional text.

Appendix

Table 13: Full prompt used for LLM-based stance quantification.

Score Range	Mapped Value	Meaning
$[0.0,0.2]$	0.1	Severe conflict in core facts or direct denial.
$(0.2,0.4]$	0.3	Certain contradictions or inconsistencies exist, with weak refutation strength.
$(0.4,0.6]$	0.5	Content is irrelevant or lacks sufficient information to support a judgment.
$(0.6,0.8]$	0.7	Possesses partial factual consistency, providing limited support.
$(0.8,1.0]$	0.9	Highly consistent at the key factual level, explicitly supporting the original claim.

Appendix

Table 14: Mapping rules for the continuous stance scores output by the LLM.

Figure 7: Distribution of external-evidence stance scores in the SSS dataset.

Figure 8: Qualitative cases comparing entity-level and event-level evidence retrieval. Event-level retrieval returns evidence that is more closely aligned with the complete news event, whereas entity-level retrieval tends to retrieve background information about salient entities mentioned in the news.

Type	Parameter	Value
Feature	Text dim.	768
	Image dim.	768
	Evidence dim.	768
	Projection dim.	768
	Relation dim. $d_{r}$	32
Graph	$\delta_{+}$	0.6

Appendix

Table 15: Detailed experimental settings.

Role	Content
System	You are a rigorous multimodal news authenticity evaluator. Your task is to determine whether a news item is fake or real by jointly considering the news text, the associated image, and the retrieved external evidence. Decision criteria: (1) Judge whether the core claim expressed by the news is supported or contradicted by the visual content and the retrieved evidence. (2) Use the retrieved evidence only as contextual evidence. Do not assume that an evidence source is always correct, and do not use prior knowledge beyond the supplied inputs. (3) If the image-text pair or the evidence reveals clear factual contradiction, temporal inconsistency, entity mismatch, or unsupported fabrication, classify the news as fake. (4) If the news text, image, and evidence are mutually consistent and the evidence supports the core claim, classify the news as real. (5) If the evidence is insufficient, make the most likely binary judgment based on the supplied news text, image, and evidence. Output requirement: Return exactly one label: Fake or Real . Do not provide explanations.
User	News text: [NEWS TEXT] News image: [NEWS IMAGE] Retrieved evidence snippets: [EVIDENCE SNIPPETS]
MLLM	Fake or Real

Appendix

Table 16: Full prompt used for direct w/ MLLM authenticity judgment.

Category	Method	Accuracy	Fake News			Real News
Category	Method	Accuracy	Precision	Recall	F1	Precision	Recall	F1
w/o EK	CAFE	0.815 $\pm 0.005$	0.850 $\pm 0.005$	0.774 $\pm 0.005$	0.810 $\pm 0.005$	0.785 $\pm 0.005$	0.858 $\pm 0.005$	0.820 $\pm 0.005$
	MRML	0.903 $\pm 0.011$	0.881 $\pm 0.010$	0.935 $\pm 0.010$	0.908 $\pm 0.010$	0.928 $\pm 0.012$	0.869 $\pm 0.011$	0.898 $\pm 0.011$
	Event-Radar	0.881 $\pm 0.010$	0.873 $\pm 0.009$	0.896 $\pm 0.010$	0.884 $\pm 0.009$	0.889 $\pm 0.010$	0.865 $\pm 0.010$	0.877 $\pm 0.010$
	MSACA	0.894 $\pm 0.006$	0.861 $\pm 0.006$	0.944 $\pm 0.005$	0.900 $\pm 0.005$	0.935 $\pm 0.005$	0.841 $\pm 0.007$	0.886 $\pm 0.006$
w/ EK	KEHGNN-FD	0.764 $\pm 0.005$	0.752 $\pm 0.004$	0.801 $\pm 0.005$	0.776 $\pm 0.004$	0.779 $\pm 0.005$	0.726 $\pm 0.005$	0.751 $\pm 0.005$

Appendix

Table 17: Overall performance on Weibo-21.

Category	Method	Accuracy	Fake News			Real News
Category	Method	Accuracy	Precision	Recall	F1	Precision	Recall	F1
w/o EK	CAFE	0.786 $\pm 0.011$	0.773 $\pm 0.009$	0.864 $\pm 0.011$	0.816 $\pm 0.010$	0.807 $\pm 0.015$	0.691 $\pm 0.012$	0.745 $\pm 0.013$
	MRML	0.860 $\pm 0.006$	0.852 $\pm 0.005$	0.900 $\pm 0.006$	0.876 $\pm 0.005$	0.870 $\pm 0.007$	0.811 $\pm 0.007$	0.839 $\pm 0.007$
	Event-Radar	0.840 $\pm 0.010$	0.831 $\pm 0.008$	0.888 $\pm 0.010$	0.859 $\pm 0.009$	0.852 $\pm 0.012$	0.780 $\pm 0.010$	0.815 $\pm 0.011$
	MSACA	0.846 $\pm 0.009$	0.839 $\pm 0.008$	0.889 $\pm 0.008$	0.864 $\pm 0.008$	0.855 $\pm 0.010$	0.794 $\pm 0.010$	0.823 $\pm 0.010$
w/ EK	KEHGNN-FD	0.763 $\pm 0.010$	0.752 $\pm 0.008$	0.846 $\pm 0.009$	0.796 $\pm 0.008$	0.780 $\pm 0.013$	0.662 $\pm 0.011$	0.716 $\pm 0.012$

Appendix

Table 18: Overall performance on Fakeddit.

Category	Method	Accuracy	Fake News			Real News
Category	Method	Accuracy	Precision	Recall	F1	Precision	Recall	F1
w/o EK	CAFE	0.726 $\pm 0.009$	0.709 $\pm 0.008$	0.802 $\pm 0.009$	0.753 $\pm 0.009$	0.750 $\pm 0.012$	0.644 $\pm 0.010$	0.693 $\pm 0.010$
	MRML	0.758 $\pm 0.006$	0.742 $\pm 0.005$	0.821 $\pm 0.006$	0.779 $\pm 0.005$	0.780 $\pm 0.007$	0.690 $\pm 0.006$	0.732 $\pm 0.006$
	Event-Radar	0.754 $\pm 0.005$	0.755 $\pm 0.005$	0.779 $\pm 0.005$	0.767 $\pm 0.005$	0.752 $\pm 0.006$	0.726 $\pm 0.005$	0.739 $\pm 0.006$
	MSACA	0.762 $\pm 0.011$	0.757 $\pm 0.010$	0.801 $\pm 0.010$	0.778 $\pm 0.010$	0.769 $\pm 0.012$	0.721 $\pm 0.012$	0.744 $\pm 0.012$
w/ EK	KEHGNN-FD	0.674 $\pm 0.010$	0.672 $\pm 0.009$	0.726 $\pm 0.010$	0.698 $\pm 0.009$	0.675 $\pm 0.011$	0.616 $\pm 0.010$	0.644 $\pm 0.011$

Appendix

Table 19: Overall performance on SSS.

Variation	Accuracy	Macro F1	Fake F1	HHF F1	HHF Recall
w/o Evidence Graph	0.913 $\pm 0.007$	0.913 $\pm 0.007$	0.915 $\pm 0.007$	0.956 $\pm 0.007$	0.961 $\pm 0.007$
– w/o Edge Weights	0.920 $\pm 0.009$	0.920 $\pm 0.009$	0.921 $\pm 0.009$	0.959 $\pm 0.009$	0.956 $\pm 0.009$
– w/o Edge Types	0.914 $\pm 0.006$	0.914 $\pm 0.006$	0.916 $\pm 0.006$	0.960 $\pm 0.006$	0.967 $\pm 0.006$
w/o Harm-aware Branch	0.915 $\pm 0.006$	0.915 $\pm 0.006$	0.914 $\pm 0.005$	0.946 $\pm 0.007$	0.919 $\pm 0.006$
– w/o $\mathcal{L}_{\mathrm{harm}}$	0.923 $\pm 0.010$	0.923 $\pm 0.010$	0.923 $\pm 0.011$	0.954 $\pm 0.010$	0.936 $\pm 0.011$
– w/o $\mathcal{L}_{\mathrm{rank}}$	0.924 $\pm 0.007$	0.924 $\pm 0.007$	0.924 $\pm 0.008$	0.947 $\pm 0.008$	0.932 $\pm 0.009$

Appendix

Table 20: Ablation results on Weibo-21.

Variation	Accuracy	Macro F1	Fake F1	HHF F1	HHF Recall
w/o Evidence Graph	0.909 $\pm 0.012$	0.907 $\pm 0.012$	0.921 $\pm 0.012$	0.939 $\pm 0.012$	0.952 $\pm 0.014$
– w/o Edge Weights	0.915 $\pm 0.004$	0.914 $\pm 0.004$	0.924 $\pm 0.004$	0.936 $\pm 0.005$	0.935 $\pm 0.005$
– w/o Edge Types	0.911 $\pm 0.011$	0.910 $\pm 0.011$	0.921 $\pm 0.011$	0.930 $\pm 0.011$	0.935 $\pm 0.011$
w/o Harm-aware Branch	0.912 $\pm 0.010$	0.911 $\pm 0.010$	0.920 $\pm 0.010$	0.925 $\pm 0.011$	0.907 $\pm 0.009$
– w/o $\mathcal{L}_{\mathrm{harm}}$	0.909 $\pm 0.010$	0.908 $\pm 0.010$	0.917 $\pm 0.010$	0.926 $\pm 0.010$	0.919 $\pm 0.009$
– w/o $\mathcal{L}_{\mathrm{rank}}$	0.906 $\pm 0.007$	0.905 $\pm 0.007$	0.913 $\pm 0.007$	0.922 $\pm 0.007$	0.900 $\pm 0.007$

Appendix

Table 21: Ablation results on Fakeddit.

Variation	Accuracy	Macro F1	Fake F1	HHF F1	HHF Recall
w/o Evidence Graph	0.822 $\pm 0.004$	0.818 $\pm 0.004$	0.844 $\pm 0.005$	0.892 $\pm 0.004$	0.956 $\pm 0.005$
– w/o Edge Weights	0.837 $\pm 0.011$	0.835 $\pm 0.010$	0.851 $\pm 0.011$	0.886 $\pm 0.010$	0.923 $\pm 0.012$
– w/o Edge Types	0.826 $\pm 0.011$	0.824 $\pm 0.011$	0.843 $\pm 0.010$	0.877 $\pm 0.011$	0.928 $\pm 0.011$
w/o Harm-aware Branch	0.839 $\pm 0.010$	0.838 $\pm 0.010$	0.850 $\pm 0.010$	0.878 $\pm 0.010$	0.905 $\pm 0.010$
– w/o $\mathcal{L}_{\mathrm{harm}}$	0.839 $\pm 0.006$	0.838 $\pm 0.006$	0.852 $\pm 0.007$	0.888 $\pm 0.007$	0.923 $\pm 0.007$
– w/o $\mathcal{L}_{\mathrm{rank}}$	0.833 $\pm 0.004$	0.832 $\pm 0.004$	0.844 $\pm 0.004$	0.892 $\pm 0.005$	0.917 $\pm 0.004$

Appendix

Table 22: Ablation results on SSS.

Category	Method	0.1	0.3	0.5	0.7	0.9	Overall
w/o EK	CAFE	0.825 $\pm 0.011$	0.806 $\pm 0.004$	0.816 $\pm 0.008$	0.805 $\pm 0.006$	0.818 $\pm 0.008$	0.815 $\pm 0.005$
	MRML	0.898 $\pm 0.011$	0.909 $\pm 0.009$	0.902 $\pm 0.012$	0.908 $\pm 0.011$	0.900 $\pm 0.008$	0.903 $\pm 0.011$
	Event-Radar	0.886 $\pm 0.012$	0.871 $\pm 0.005$	0.882 $\pm 0.010$	0.879 $\pm 0.012$	0.884 $\pm 0.008$	0.881 $\pm 0.010$
	MSACA	0.889 $\pm 0.009$	0.898 $\pm 0.007$	0.892 $\pm 0.005$	0.900 $\pm 0.006$	0.890 $\pm 0.001$	0.894 $\pm 0.006$
w/ EK	KEHGNN-FD	0.771 $\pm 0.009$	0.756 $\pm 0.005$	0.769 $\pm 0.008$	0.755 $\pm 0.007$	0.765 $\pm 0.002$	0.764 $\pm 0.005$
	NSLM	0.882 $\pm 0.009$	0.871 $\pm 0.010$	0.879 $\pm 0.010$	0.876 $\pm 0.009$	0.881 $\pm 0.009$	0.878 $\pm 0.009$

Appendix

Table 23: Accuracy across harm levels on Weibo-21.

Category	Method	0.1	0.3	0.5	0.7	0.9	Overall
w/o EK	CAFE	0.792 $\pm 0.011$	0.782 $\pm 0.012$	0.789 $\pm 0.011$	0.783 $\pm 0.012$	0.786 $\pm 0.010$	0.786 $\pm 0.011$
	MRML	0.852 $\pm 0.008$	0.865 $\pm 0.009$	0.858 $\pm 0.009$	0.863 $\pm 0.002$	0.860 $\pm 0.001$	0.860 $\pm 0.006$
	Event-Radar	0.844 $\pm 0.008$	0.836 $\pm 0.011$	0.842 $\pm 0.011$	0.837 $\pm 0.011$	0.839 $\pm 0.006$	0.840 $\pm 0.010$
	MSACA	0.840 $\pm 0.010$	0.849 $\pm 0.008$	0.844 $\pm 0.011$	0.850 $\pm 0.011$	0.847 $\pm 0.001$	0.846 $\pm 0.009$
w/ EK	KEHGNN-FD	0.765 $\pm 0.007$	0.758 $\pm 0.011$	0.766 $\pm 0.012$	0.761 $\pm 0.012$	0.763 $\pm 0.006$	0.763 $\pm 0.010$
	NSLM	0.844 $\pm 0.011$	0.855 $\pm 0.010$	0.847 $\pm 0.006$	0.854 $\pm 0.011$	0.849 $\pm 0.005$	0.850 $\pm 0.006$

Appendix

Table 24: Accuracy across harm levels on Fakeddit.

Category	Method	0.1	0.3	0.5	0.7	0.9	Overall
w/o EK	CAFE	0.731 $\pm 0.009$	0.722 $\pm 0.010$	0.729 $\pm 0.011$	0.724 $\pm 0.010$	0.722 $\pm 0.007$	0.726 $\pm 0.009$
	MRML	0.752 $\pm 0.005$	0.762 $\pm 0.006$	0.756 $\pm 0.006$	0.761 $\pm 0.006$	0.756 $\pm 0.005$	0.758 $\pm 0.006$
	Event-Radar	0.755 $\pm 0.010$	0.748 $\pm 0.011$	0.760 $\pm 0.002$	0.751 $\pm 0.012$	0.752 $\pm 0.006$	0.754 $\pm 0.005$
	MSACA	0.757 $\pm 0.012$	0.768 $\pm 0.011$	0.760 $\pm 0.012$	0.765 $\pm 0.011$	0.759 $\pm 0.009$	0.762 $\pm 0.011$
w/ EK	KEHGNN-FD	0.680 $\pm 0.010$	0.668 $\pm 0.011$	0.677 $\pm 0.010$	0.671 $\pm 0.011$	0.674 $\pm 0.008$	0.674 $\pm 0.010$
	NSLM	0.775 $\pm 0.010$	0.785 $\pm 0.010$	0.778 $\pm 0.011$	0.784 $\pm 0.011$	0.780 $\pm 0.012$	0.781 $\pm 0.011$

Appendix

Table 25: Accuracy across harm levels on SSS.

Dataset	Predictor	MAE	RMSE
Weibo-21	Text-only MLP	0.103 $\pm 0.005$	0.137 $\pm 0.005$
Weibo-21	RAEGNet harm branch	0.076 $\pm 0.005$	0.103 $\pm 0.005$
Fakeddit	Text-only MLP	0.112 $\pm 0.005$	0.145 $\pm 0.004$
Fakeddit	RAEGNet harm branch	0.084 $\pm 0.006$	0.112 $\pm 0.006$
SSS	Text-only MLP	0.119 $\pm 0.008$	0.153 $\pm 0.008$
SSS	RAEGNet harm branch	0.091 $\pm 0.007$	0.119 $\pm 0.006$

Appendix

Table 26: Direct evaluation of conditional-harm prediction.

Dataset	Evaluation Set	Model	Accuracy	Macro F1	Fake F1	HHF F1
Weibo-21	Original	w/o Harm-aware Risk	0.915 $\pm 0.006$	0.915 $\pm 0.006$	0.914 $\pm 0.005$	0.946 $\pm 0.007$
	Original	Full Model	0.931 $\pm 0.009$	0.931 $\pm 0.009$	0.934 $\pm 0.009$	0.972 $\pm 0.006$
	Harm-matched	w/o Harm-aware Risk	0.902 $\pm 0.010$	0.902 $\pm 0.009$	0.903 $\pm 0.010$	0.928 $\pm 0.010$
	Harm-matched	Full Model	0.914 $\pm 0.010$	0.914 $\pm 0.010$	0.916 $\pm 0.010$	0.946 $\pm 0.010$
Fakeddit	Original	w/o Harm-aware Risk	0.912 $\pm 0.010$	0.911 $\pm 0.010$	0.920 $\pm 0.010$	0.925 $\pm 0.011$
	Original	Full Model	0.919 $\pm 0.007$	0.918 $\pm 0.007$	0.926 $\pm 0.006$	0.947 $\pm 0.006$

Appendix

Table 27: Harm-matched evaluation. Held-out samples are balanced between real and fake news within each harm level in every fold.

Dataset	Retrieval Query	Backbone	Accuracy	Macro F1	Fake F1	HHF F1
Weibo-21	Entity-level	Evidence Fusion	0.889 $\pm 0.004$	0.889 $\pm 0.004$	0.891 $\pm 0.004$	0.932 $\pm 0.005$
	Entity-level	RAEGNet	0.914 $\pm 0.005$	0.914 $\pm 0.005$	0.916 $\pm 0.005$	0.955 $\pm 0.006$
	Event-level	Evidence Fusion	0.907 $\pm 0.010$	0.907 $\pm 0.010$	0.910 $\pm 0.010$	0.946 $\pm 0.010$
	Event-level	RAEGNet	0.928 $\pm 0.005$	0.928 $\pm 0.005$	0.930 $\pm 0.004$	0.968 $\pm 0.003$
Fakeddit	Entity-level	Evidence Fusion	0.872 $\pm 0.008$	0.870 $\pm 0.008$	0.884 $\pm 0.009$	0.912 $\pm 0.009$
	Entity-level	RAEGNet	0.899 $\pm 0.012$	0.898 $\pm 0.011$	0.908 $\pm 0.012$	0.934 $\pm 0.011$

Appendix

Table 28: Controlled comparison of retrieval granularity and model architecture under the same source pool, filtering process, and evidence budget.

Dataset	Method	Evidence Source	Accuracy	Fake F1	Real F1
Weibo-21	KEHGNN-FD	Original	0.764 $\pm 0.005$	0.776 $\pm 0.004$	0.751 $\pm 0.005$
	KEHGNN-FD	ELERF	0.801 $\pm 0.008$	0.812 $\pm 0.008$	0.789 $\pm 0.008$
	NSLM	Original	0.878 $\pm 0.009$	0.885 $\pm 0.008$	0.870 $\pm 0.010$
	NSLM	ELERF	0.896 $\pm 0.005$	0.901 $\pm 0.005$	0.890 $\pm 0.004$
	ERIC-FND	Original	0.844 $\pm 0.011$	0.857 $\pm 0.010$	0.828 $\pm 0.012$
	ERIC-FND	ELERF	0.879 $\pm 0.011$	0.886 $\pm 0.011$	0.871 $\pm 0.010$

Appendix

Table 29: Effect of replacing the original external evidence of w/ EK baselines with ELERF.

Explore similar work

Apr 20, 2026cs.CL

Retrieval-Augmented Multimodal Model for Fake News Detection

In recent years, multimodal multidomain fake news detection has garnered increasing attention. Nevertheless, this direction presents two significant challenges: (1) Failure to Capture Cross-Instance Narrative Consistency: existing models usually evaluate each news in isolation, fail to capture cross-instance narrative consistency, and thus struggle to address the spread of cluster based fake news driven by social media; (2) Lack of Domain Specific Knowledge for Reasoning: conventional models, which rely solely on knowledge encoded in their parameters during training, struggle to generalize to new or data-scarce domains (e.g., emerging events or niche topics). To tackle these challenges, we introduce Retrieval-Augmented Multimodal Model for Fake News Detection (RAMM). First, RAMM employs a Multimodal Large Language Model (MLLM) as its backbone to capture cross-modal semantic information from news samples. Second, RAMM incorporates an Abstract Narrative Alignment Module. This component adaptively extracts abstract narrative consistency from diverse instances across distinct domains, aggregates relevant knowledge, and thereby enables the modeling of high-level narrative information. Finally, RAMM introduces a Semantic Representation Alignment Module, which aligns the model's decision-making paradigm with that of humans - specifically, it shifts the model's reasoning process from direct inference on multimodal features to an instance-based analogical reasoning process. Extensive experimental results on three public datasets validate the efficacy of our proposed approach. Our code is available at the following link: https://github.com/li-yiheng/RAMM

Yiheng Li, Weihai Lu, Hanyi Yu +1

University of International Business and Economics · Beijing, China · Peking University +4

Fake News Detection Fake News Multimodal Large Language Models Cross-Modal

Ask PDF

Aug 2, 2026cs.AI

MRAFnd: Multimodal Retrieval-Augmented Framework for Zero-Shot Fake News Detection

The rapid dissemination of multimodal content has intensified the spread of fabricated news, presenting a substantial threat to social integrity. A formidable challenge for current detection systems is identifying misinformation related to novel events in zero-shot scenarios. Prevailing zero-shot methods typically assess news items in isolation via semantic matching, a strategy that fails to recognize the recycled disinformation tactics from past campaigns and lacks the sophisticated reasoning needed to identify subtle, cross-modal discrepancies. To surmount these deficiencies, we introduce \textbf{MRAFnd}, a novel \underline{\textbf{M}}ultimodal \underline{\textbf{R}}etrieval-\underline{\textbf{A}}ugmented Framework for Zero-Shot \underline{\textbf{F}}ake \underline{\textbf{N}}ews \underline{\textbf{D}}etection. MRAFnd emulates a collaborative team of analysts to verify news veracity. The framework initiates with \textbf{Multimodal Similarity-based News Retrieval} to assemble a corpus of contextually analogous articles from an unlabeled reference database. Subsequently, during the \textbf{Bifurcated Evidential Reasoning} stage, agents perform a dual-directional analysis to extract critical patterns from the retrieved evidence. Finally, a \textbf{Multi-Agent Collaborative Debate}, involving Analyst and Arbiter agents, engages in a structured discourse to arrive at a definitive and robust conclusion. Comprehensive experiments on three benchmark datasets reveal that MRAFnd markedly surpasses state-of-the-art baselines, achieving an accuracy gain of up to 2.35% on the demanding Weibo-21 dataset.

Lehan Zhang, Yinlei Cheng, Shiqi Hu Yiheng Zhou +2

Beijing Institute of Fashion Technology, Beijing, China · Shenyang University of Technology, Shenyang, China

Fake News Detection Reference-Guided Multimodal In-Context Verification Zero-Shot Robustness Multimodal Retrieval Multi-Agent Debate

Ask PDF

Sep 29, 2026cs.CL

Rethinking Multimodal Fake News Detection in the Generative AI Era

Generative content is increasingly entering the production and dissemination of news, transforming fake news from manually fabricated or simply manipulated material into complex forms in which native and generated content jointly participate. Existing multimodal fake news detection research primarily focuses on veracity assessment and rarely characterizes how generativity differences affect the reliability of evidence. In contrast, AIGC detection primarily determines whether content is generated or modified by generative models, but it does not by itself establish whether the underlying news event is true. To bridge the separation between these tasks in data and evaluation, we construct Weibo26, a multimodal fake news detection dataset for generative-content scenarios. On this basis, we propose the Generativity-Aware Hierarchical Reasoning (GAHR) framework, which combines global judgment with local correction so that generativity information participates in news-veracity reasoning. Experiments on multiple existing fake news detection benchmarks and Weibo26 show that GAHR achieves competitive veracity-detection performance while effectively identifying generative content.

Wenbin Shen, Guoxuan Qin, Guangxu Yao +3

Nanjing University of Science and Technology, Nanjing, China · University of Chinese Academy of Sciences, Beijing, China

Fake News Detection Ai-Generated Content Detection Fake News Generative Artificial Intelligence

Ask PDF