LM-CAGF: A prompt-enhanced Cross-Attention and Gated Fusion framework for multimodal fake news detection

Abstract

Background and Aim: The rapid proliferation of misinformation, particularly fake news on social media, poses a critical challenge to information credibility and public trust. However, existing fake news detection methods face two major limitations: they often rely on lightweight or task-specific encoders that fail to capture rich semantics, and their fusion strategies are typically shallow, overlooking fine-grained inconsistencies between text and images. Methodology: To overcome these limitations, we propose LM-CAGF (Large-Model-based Cross-Attention and Gated Fusion), a novel prompt-enhanced framework for multimodal fake news detection. LM-CAGF leverages large language models (LLMs) for textual encoding and large-scale vision–language pretrained models (VL-PLMs) for visual representation, which are aligned through a bidirectional cross-attention mechanism and integrated via a gated fusion module. The fused representation is further embedded into task-specific prompts and processed by a pretrained language model to enable hypothesis-driven reasoning. Results: Extensive experiments on three benchmark datasets demonstrate that LM-CAGF achieves competitive performance and generally improves over most baseline methods. In addition, ablation studies confirm the contributions of cross-attention, gated fusion, and verbalizer design, highlighting the effectiveness of large-model-driven prompt learning for robust multimodal fake news detection. Conclusion: Overall, LM-CAGF provides an effective prompt-enhanced multimodal framework that improves cross-modal semantic alignment and enhances detection performance in multimodal fake news detection tasks.