RumorGAL: Graph-anchored LLM for domain-generalized rumor detection

Abstract

Rumor detection models must remain robust to breaking events that are unseen during training. Although large language models (LLMs) offer strong generalization potential, their text-centric nature makes them vulnerable to spurious textual patterns in noisy social media discussions and leaves them without explicit awareness of how rumors propagate through conversation threads. To address these limitations, we propose RumorGAL, a graph-anchored LLM framework for domain-generalized rumor detection. The framework first employs a graph neural network (GNN) to derive graph-anchored structural prefixes from the propagation graph, providing structural guidance for the LLM. It then identifies high-quality textual evidence from noisy discussions through an evidence density-guided mining strategy. Finally, a hierarchical reasoning mechanism organizes both the structural prefixes and textual evidence into a unified reasoning process, enabling the LLM to ground its predictions in both propagation patterns and informative content. Across the three transfer settings, RumorGAL achieves robust performance, improving average accuracy over the strongest baseline, zero-shot GPT-4.1, by 5.27 percentage points under the domain generalization setting. When compared with domain adaptation methods, it surpasses the best domain adaptation method in two out of three tasks, with improvements of 3.1 and 0.5 percentage points in F 1 score for the rumor class, respectively. These findings suggest that an LLM-as-primary, GNN-as-auxiliary design, which couples structure-aware guidance with evidence-grounded reasoning, is a promising approach to robust and interpretable rumor detection.