Title: From Zero to Hero: Advancing Zero-Shot Foundation Models for Tabular Outlier Detection

Original: arXiv:2602.03018v2 Announce Type: replace

Abstract:

While outlier detection (OD) is a staple in practical applications, its successful adaptation to novel tasks is often impeded by the scarcity of labeled outliers. This data gap complicates the notoriously difficult process of selecting appropriate algorithms and hyperparameters. Although foundation models (FMs) have revolutionized machine learning, OD has seen similar transformative potential, as demonstrated by Shen et al. (2025) with the introduction of FoMo-0D, the pioneering FM for OD that outperformed many existing baselines.

Building on this progress, we present OUTFORMER, an enhanced framework that improves upon FoMo-0D through two key innovations: a mixture of synthetic priors and self-evolving curriculum training. OUTFORMER is pretrained exclusively on synthetic labeled datasets and leverages in-context learning by treating the training data of a new task as input to infer test labels. This approach enables fast, zero-shot inference, relying solely on a forward pass without the need for labeled outliers. By eliminating the requirement for additional model training or specialized model selection, OUTFORMER facilitates a truly plug-and-play deployment. The model establishes state-of-the-art results on the prominent AdBench dataset, as well as on two new large-scale OD benchmarks we introduced, which encompass more than 1,500 datasets, all while maintaining rapid inference speeds.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC