PFN features power tabular anomaly detection: ZEN tops ADBench without fine-tuning
Maximilian Bershtman · hf · 2026-10-06
This work adapts prior-data fitted network (PFN) representations for tabular anomaly detection, where no labeled anomalies exist before deployment and the reference set may itself contain anomalies. The authors first use frozen TabPFN features with nearest-neighbor distance scoring, identifying the best layers and extraction procedure. They then fine-tune the model on the reference set to better separate normal from anomalous samples. On ADBench, the fine-tuning-free method ZEN achieves a higher mean AUROC than every baseline, the fine-tuned FOCUS improves further, and the approach generalizes across PFN models.
More from Research
- Researcher unveils RSI paradigm: generic disposable agents plus an evolving knowledge base — yisongyue · 2026-10-07
- KNOWS Benchmark Jointly Evaluates Agents on Search, Tools and Visual Understanding — anmarasovic · 2026-10-07
- ID-Forcing Keeps Long Video Generation In-Distribution, Enabling Minute-Scale Videos Without Fine-Tuning — _akhaliq · 2026-10-07
- Harness-Aware Distillation Boosts Small-Model Agent Harness Use 65.7% to 73.1% on ALFWorld — omarsar0 · 2026-10-07
- Irregular launches containment challenges to stress-test evaluation security — evijit · 2026-10-07
- Priced Guidance: quantifying whether LLMs can generate novel research ideas via compression — ChengleiSi · 2026-10-07