Adapting prior-data fitted networks for tabular anomaly detection
PFN features are being tested as a practical representation layer for tabular anomaly detection.
The paper adapts TabPFN-style representations to a setting where anomalies are not labeled before deployment. A frozen-feature method, ZEN, scores samples by nearest-neighbor distance in feature space and reports stronger mean AUROC than baselines on ADBench. A fine-tuned version, FOCUS, uses the reference set to further separate normal samples from anomalies. The authors say the approach generalizes across PFN models. HF Daily Papers' note
The paper adapts TabPFN-style representations to a setting where anomalies are not labeled before deployment. A frozen-feature method, ZEN, scores samples by nearest-neighbor distance in feature space and reports stronger mean AUROC than baselines on ADBench. A fine-tuned version, FOCUS, uses the reference set to further separate normal samples from anomalies. The authors say the approach generalizes across PFN models. HF Daily Papers' note
score 4