Aug 2026· Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2· pp. 461-472· 0 citations· 22 references
Abstract
Multimodal fake news combining text and images has become increasingly prevalent, fueled by the rapid dissemination on social media. Existing approaches predominantly rely on supervised learning–driven small multimodal language models, yet they are constrained by the knowledge scope and logical reasoning capabilities limited by training data: their performance degrades significantly when encountering novel content not covered in the training data, and they typically provide only uninterpretable binary true/false predictions. To address these challenges, we propose the Metacognitive Agentic Reasoning for Multimodal Fake News Detection(MAR), a framework that integrates multi-agent metacognitive debate with external knowledge retrieval to enhance the accuracy, generalization, and interpretability of multimodal fake news detection. At its core, MAR introduces a news-domain informed and metacognition-inspired multi-agent reasoning mechanism: it first generates several prior pseudo-labels of the news domain, and defines the characteristics of two opposing agents (i.e., Believer and Skeptic) leveraging these pseudo-labels; then, the Believer and Skeptic go through a three-stage ''Draft – Self-Critique – Refine'' interactive debate simulating humans' learning behavior, called the metacognitive debate. Specifically, MAR first generates an initial evidence-grounded judgment (Draft), then critically reflects on their own and others' arguments (Self-Critique), and finally iteratively refines their conclusions by incorporating internal or external evidence (Refine). To improve factual reliability, the framework dynamically retrieves trustworthy external knowledge via web and reverse image searches, thereby mitigating the hallucinations inherent in large language models. Experiments show that MAR achieves state-of-the-art performance on two benchmarks and significantly outperforms existing methods in terms of generalization and interpretability. The source code is available at https://github.com/Averdgr/MAR_KDD.
The proliferation of social media has created fertile ground for misinformation, a challenge further intensified by recent advances in generative artificial intelligence. Modern fake news increasingly takes the form of sophisticated multimodal campaigns, where synthetic images and stylistically manipulated text are jointly employed to evade existing detection systems. Despite substantial progress, real-world deployment of multimodal fake news detection models remains constrained by an ''impossible triangle'' of accuracy, inference efficiency, and robustness. To address these challenges, we propose DAR-Lite, a serial two-stage framework that rethinks the detection pipeline through explicit decoupling of representation denoising and contextual reasoning. In the first stage, a variational information bottleneck distills compact and noise-invariant semantic anchors from raw image--text inputs, reducing sensitivity to nuisance factors. In the second stage, an adaptive reasoning engine integrates retrieval-augmented verification, social credibility propagation, and dynamic propagation signals via gated cross-attention to perform structured reasoning over heterogeneous contexts. An auxiliary logical fallacy detection task further encourages reasoning beyond surface-level pattern matching. Extensive experiments on multiple large-scale benchmarks demonstrate that DAR-Lite consistently outperforms state-of-the-art methods, particularly under low false-positive constraints critical for real-world applications. Efficiency analyses further show that the proposed serial architecture achieves a favorable balance between detection performance and computational cost, making DAR-Lite suitable for practical, large-scale misinformation detection.
Maolin Wang, Ziting Mai, Zichun Liu et al.· Proceedings of the 32nd ACM...· 0 citations
The results indicate that explicitly modeling semantic conflict as a discriminative feature effectively improves detection precision and generalization, providing a robust solution for factual verification in complex media environments.
Zi-Heng Wang, Junfang Song, Shuyu Wang et al.· Multimedia Systems· 0 citations
In particular, the rapid dissemination of fake news using social media platforms has posed a severe threat to social stability and information reliability in language groups that lack sufficient resources. In addition, it has become difficult to automatically identify fake news in the case of Bengali digital media because fake news is mostly distributed using multimodal data like memes, screenshots, posters, photos with texts, etc. The effectiveness of the existing fake news detection methods is somewhat hindered by their inability to provide explainability and their focus mainly on either textual or visual data. This study proposes a context-aware multimodal reasoning approach for explainable Bengali fake news detection. The proposed model incorporates EasyOCR for Bengali textual information extraction, ResNet-50 and ViT for supplementing visual feature learning, and Qwen2-VL-2B-Instruct for multimodal semantic inference. The proposed approach can detect semantic contradictions and associations between textual assertions and image information through matching textual and visual data based on a context-sensitive fusion technique. This methodology generates interpretable explanations beyond the fake/real binary categorization to enhance user trust in automated outcomes. According to an experiment conducted using a multimodal Bengali fake news dataset, the proposed approach surpasses conventional CNN, transformer, and unimodal baselines on several performance metrics. Results indicate how context-based multimodal reasoning can improve the efficiency and robustness of the model, along with making it more interpretable. The proposed method serves as an encouraging route towards the identification of fake news in multiple low resource languages, as well as in combating misinformation in the Bengali digital ecosystem.
The proliferation of Large Language Models (LLMs) has significantly lowered the barriers to generating sophisticated misinformation, posing unprecedented challenges to fake news detection systems. Although generative agents have shown promise in constructive applications like fact-checking workflows, their potential for collaborative fake news generation remains unexplored. This paper introduces the first multi-agent framework for fake news generation, where specialized LLM-powered agents collaboratively transform authentic news articles into falsified yet plausible versions. Our pipeline employs five specialized agents: a Semantic Analyzer that extracts information-dense components, a Salient Sentence Editor and Number Modifier that manipulate key textual and numerical elements, a Narrative Modifier that ensures overall coherence, and a Title Editor that generates contextually appropriate headlines. We evaluate this system on two benchmark datasets, the ISOT Fake News Dataset and the All The News Dataset, against six state-of-the-art detection models. Our results demonstrate that the proposed framework achieves deception success rates exceeding 90% while maintaining high semantic similarity to original articles. Notably, even smaller generative models can surpass 80% success against high-capacity detectors, highlighting critical vulnerabilities in current fake news detection systems and underscoring the urgent need for more robust defense mechanisms against coordinated generative AI attacks.
Gian Marco Orlando, Diego Russo, Valerio La Gatta et al.· ACM Transactions on Intellig...· 0 citations
Readers of online news often lack the time and domain expertise required to verify unfamiliar claims and sources. Professional fact-checkers address this gap through lateral reading, an iterative workflow of asking investigative questions, searching for external evidence, and synthesizing findings with attribution. We present an iterative multi-agent Retrieval-Augmented Generation (RAG) system that operationalizes this workflow for the TREC 2025 DRAGUN Track. Given a news article, specialized agents (1) generate investigative queries, (2) retrieve and filter evidence from the MS MARCO V2.1 Segmented Corpus using a three-stage retriever (BM25+RM3, cross-encoder reranking, and LLM-based selection), and (3) apply an information-sufficiency evaluator that decides whether additional searching is required before writing. The final report generator produces a 250-word trustworthiness report grounded in retrieved segments, guided by automatically generated critical investigative questions. On the official DRAGUN rubric-based evaluation with 30 news articles, our system using GPT-4.1 ranked first on report generation quality, achieving the highest mean supportive score (0.230) with low contradiction (0.013).
Dake Zhang, Mark D. Smucker· Annual International ACM SIG...· 1 citation
An Adaptive Binary Osprey Optimization Algorithm and Cross-modal Disentangled Fusion model (ABOOA-CDF) is proposed that effectively improves detection performance and effectiveness in feature optimization, cross-modal relation modeling, and semantic fusion.
