Detecting implicit inconsistency in multimodal fake news via fine-grained cross-modal adaptive reasoning

In multimodal fake news detection, identifying implicit inconsistency, where falsified details are hidden within seemingly coherent contexts, remains a critical challenge. High-level thematic consistency often masks fine-grained logical conflicts, rendering traditional global fusion methods ineffective. To address this, we propose FG-CAR, a Fine-Grained Cross-modal Adaptive Reasoning framework. FG-CAR decomposes multimodal posts into syntax-aware textual entities and salient visual regions, and constructs a multi-relation cross-modal reasoning graph with semantic alignment, attribute verification, and intra-modal structural support edges. To improve robustness under scarce implicit-inconsistency samples, we introduce a hierarchical multi-granularity contrastive learning mechanism that jointly preserves micro-level structural invariance and macro-level semantic consistency. Furthermore, a Conflict-aware Adaptive Fusion module (CAF) dynamically balances global semantic alignment and fine-grained graph reasoning according to cross-modal discrepancy. Extensive experiments on multiple public benchmark datasets demonstrate that FG-CAR consistently outperforms state-of-the-art baselines. Ablation studies and robustness analyses further verify the effectiveness of the proposed reasoning graph, contrastive learning strategy, and adaptive fusion mechanism. The source code is available at https://github.com/Wz-f/FG-CAR .