AI-generated text detection on HC3 cross-benchmark transfer
99.8AUROC (Finance)FOMAML+LoRA
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| FOMAML+LoRAAdaptation protocol=K=10-shot adaptation, Source Training Dataset=MAGE2026.07 | 99.8 | 92.2 | 99.1 | 81.3 | 96.4 | 64.2 | 99.8 | 96.6 | 99.8 | 95.8 | 98.9 | 86 | |
| Vanilla RoBERTa baselineAdaptation protocol=Zero-shot, Source Training Dataset=MAGE, Backbone=RoBERTa2026.07 | 99.7 | 97.2 | 97.6 | 82 | 97.7 | 88.8 | 99.7 | 92.4 | 99.2 | 87.5 | 98.8 | 89.6 | |
| confidence-weighted ensembleAdaptation protocol=Ensemble of Zero-shot and K-shot, Source Training Dataset=MAGE2026.07 | 99.7 | 97.2 | 97.9 | 88 | 97.8 | 89.4 | 99.7 | 96.8 | 99.5 | 89.3 | 98.9 | 92.1 |