Domain-Adaptive Multimodal Fake News Detection via Progressive Reasoning with Large Language Models

Multimodal fake news detection requires identifying semantic conflicts between text and images while accounting for differences across news domains. Existing approaches often provide limited analysis of claim–evidence relationships and view reliability. Using large language model (LLM) reasoning, we propose Progressive Domain-Adaptive Multi-view Fake News Detection with LLM Reasoning (ProDFND-LR). Its Reasoning-enhanced Multi-view Perception module combines frozen pretrained models, domain-specific low-rank adapters, and view-specific rationales to extract textual, visual, and cross-modal features. Evidence Chain Structured Reasoning links claims to evidence locations and assigns entail, contradict, or unknown relations. Counterfactual-Driven Reliability Routing verifies evidence through interventions and multiple reasoning paths, then adjusts view weights using domain information and verification results. A Progressive Reasoning Curriculum organizes training into unimodal warm-up, cross-modal alignment, and global fusion optimization. Experiments yield macro-F1 scores of 0.951 and 0.959 on Weibo and Weibo-21, respectively, and an F1-Fake score of 0.776 on FineFake, consistently outperforming the evaluated multi-domain, explicit-reasoning, and LLM-LoRA baselines. On the VeriTaS subset and X-POSE, ProDFND-LR achieves macro-F1 scores of 0.725 and 0.707, consistently outperforming all evaluated input-only baselines without relying on external web retrieval. Ablation results indicate that view decomposition, structured evidence, and reliability verification contribute to detection performance.