A Serial Two-Stage Framework for Robust Multimodal Fake News Detection via Adaptive Reasoning

Abstract

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.