GGADBench: A Comprehensive Benchmark for Generalist Graph Anomaly Detection

Generalist Graph Anomaly Detection (GGAD) aims to detect anomalies across different graphs with a unified model and has developed rapidly in recent years. However, existing methods are evaluated on different datasets and under different transfer settings and protocols, making their performance difficult to compare reliably. We present GGADBench, the first comprehensive benchmark for GGAD, covering 30 methods and 22 graphs with both organic and injected anomalies under a unified evaluation protocol. GGADBench evaluates methods in terms of overall effectiveness, cross-graph transferability, anomaly retrieval, and efficiency. We further introduce anomaly residual polarity, a dataset-level measure that characterizes how anomalies differ from their local neighborhoods. Our evaluation reveals three key findings: (i) prompt-tuned graph foundation models are competitive with specialized GGAD methods on organic targets, whereas strong AUROC does not necessarily translate into effective anomaly retrieval; (ii) cross-graph performance is driven more by the target graph than by the source anomaly type or source quantity; and (iii) residual-based rankings can reverse on targets with negative ARP. These findings highlight the importance of transferable feature construction and target-adaptive anomaly modeling for future GGAD.