MLCL: LVLM-guided multi-level contrastive learning for fake news detection

The widespread dissemination of multi-modal disinformation on social media platforms poses a serious threat to public safety and social stability, elevating the need for automated detection of such content. Although existing studies have started to incorporate multimodal data, effectively modeling cross-modal interactions remains difficult, often leading to insufficient mining of information within individual modalities. To address this issue, this paper proposes a novel LVLM-Guided multi-level contrastive learning network (MLCL). Within the textual modality, BERT and CLIP encoders extract features at varying granularities, while ResNet and CLIP encoders capture visual representations. Additionally, a Large Vision-Language Model is leveraged to derive text-based representations that reflect both local details and global contexts from images. We further introduce a multi-level alignment strategy: inside each modality, a normalized InfoNCE loss aligns features from the two encoders to reduce redundancy. Before cross-modal fusion, we jointly align the three modality-level features using both InfoNCE loss and batch-hard triplet loss, thereby explicitly modeling their consistency and relative distance. A modality-wise attention module is also incorporated to adaptively reweight and aggregate features. Experiments on two real-world datasets show that our approach outperforms existing methods, achieving overall accuracy improvements of 2.2% on Weibo and 1.5% on Gossipcop. The source code is publicly available at https://github.com/Ylh-s/MLCL .