Beyond semantics: Exploiting propagation structures with dual-adapter LLMs for fake news detection

The rapid spread of fake news across multiple platforms and domains poses significant risks to public opinion and societal stability. Traditional approaches to fake news detection, which predominantly rely on content analysis, are becoming less effective as large language models (LLMs) lower the cost of generating fake news. It is necessary to explore more robust, generalized detection methods that can handle complex real-world situations. This paper proposes Propagation Structure-Augmented LLMs (PSALLM), a novel fake news detection model that leverages LLMs to understand and perceive news propagation structures. Specifically, we design two complementary adapters within a pre-trained LLM model: an adapter with non-weight-decay (ANWD) to capture invariant structural features across domains, and an adapter with weight-decay (AWD) to capture domain-specific structural shifts. We further introduce an adaptive distillation strategy based on domain similarity to dynamically balance the influence of both adapters. Extensive experiments on four real-world cross-domain datasets demonstrate that PSALLM consistently outperforms state-of-the-art baselines, confirming its effectiveness in handling diverse news domains and manipulation strategies. The code will be published after the end of the anonymity period.