AnTAGAD: Generalist Text-Attributed Graph Anomaly Detection with Textual and Topological Anchors

Representing the rich textual semantics of text-attributed graphs (TAGs) with large language models markedly improves cross-domain generalization, enabling effective anomaly detection across domains as varied as social platforms, financial transaction systems, and multi-agent collaboration environments. Concretely, this generalizability allows detectors to recall a large fraction of anomalies across domains, as reflected in high AUROC. However, this apparent success is undermined by persistently low AUPRC, meaning many flagged nodes are false positives and individual predictions carry unreliable confidence, leading to unacceptable misjudgments in practical deployment. We trace this failure to the widespread reliance on generalizable prototype representations, which enhance cross-domain adaptation at the expense of instance-specific detail and thus blur the distinctions among individual anomalies. We term this the Over-Generalization ($\mathcal{OG}$) problem. To address this, we propose a novel multi-anchor representation scheme to replace prototypes, encoding each node against a set of high-confidence anomaly and normal anchors to improve generalization while preserving the instance-specific characteristics of individual anomalies. Extensive experiments on eight real-world cross-domain datasets show that our method consistently outperforms all competitive baselines, with substantial gains in both AUPRC and AUROC. Beyond these empirical gains, we provide a theoretical characterization of sufficient conditions for controlling $\mathcal{OG}$ together with matched-coverage empirical analyses showing reduced anomaly acceptance on most evaluated settings.