SAGAD: A Graph Foundation Model for Generalist Few-Shot Anomaly Detection via Fine-Grained Disentanglement

While graph foundation models (GFMs) have achieved remarkable success in various downstream tasks, generalizing them to graph anomaly detection (GAD) remains a formidable challenge. A critical bottleneck is that existing methods treat anomalies as a monolithic concept, overlooking the intrinsic heterogeneity between attribute conflicts and structural violations. This entangled modeling leads to insufficient fine-grained perception capabilities, especially when transferring across diverse datasets in few-shot scenarios. To address this, we propose structural-anomaly perception and attribute-anomaly perception GAD (SAGAD), a novel GAD-oriented GFM tailored for fine-grained anomaly perception. Driven by the insight that anomaly generation mechanisms remain invariant across diverse graphs, SAGAD disentangles heterogeneous patterns into a unified residual space via a dual-channel framework. In particular, we design an attribute channel utilizing self-neighbor disentanglement (SND) to isolate semantic deviations, and a structural channel employing node-community analysis to quantify topological irregularities. It then aligns these residuals with graph-agnostic dual prototypes and leverages lightweight prompt fine-tuning for rapid downstream adaptation. Comprehensive experiments on nine real-world datasets demonstrate that SAGAD achieves highly competitive performance compared to state-of-the-art methods under the GAD-oriented GFMs paradigm in few-shot scenarios.