Adapting prior-data fitted networks for tabular anomaly detection
Frozen TabPFN features alone beat the listed anomaly-detection baselines on ADBench by mean AUROC.
Bershtman and Cohen test prior-data fitted network representations for tabular anomaly detection, where no labeled anomalies are available before deployment. Their ZEN method scores samples by nearest-neighbor distance in frozen TabPFN feature space after choosing task-suited layers and extraction steps. They then fine-tune on the reference set with FOCUS, reporting further gains despite possible contamination by anomalies. The paper says the approach also carries across PFN models. ArXiv · AI/CL/LG's note
Bershtman and Cohen test prior-data fitted network representations for tabular anomaly detection, where no labeled anomalies are available before deployment. Their ZEN method scores samples by nearest-neighbor distance in frozen TabPFN feature space after choosing task-suited layers and extraction steps. They then fine-tune on the reference set with FOCUS, reporting further gains despite possible contamination by anomalies. The paper says the approach also carries across PFN models. ArXiv · AI/CL/LG's note
score 4