Near-OOD Detection on AGNews one-vs-all runs
0.9AUROC (c3)Gemma2-2b + 16k SAE + Markov
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Gemma2-2b + 16k SAE + MarkovBackbone=Gemma2-2b, SAE Configuration=16k, Detection Backend=Markov, Evaluation Protocol=ID-only training2026.05 | 0.9 | 0.98 | 0.92 | 0.88 | 0.96 | 0.99 | 0.97 | 0.95 | 0.5 | 0.03 | 0.41 | 0.51 | |
| Gemma2-9b + 16k SAE + MarkovBackbone=Gemma2-9b, SAE Configuration=16k, Detection Backend=Markov, Evaluation Protocol=ID-only training2026.05 | 0.87 | 0.9 | 0.83 | 0.81 | 0.95 | 0.96 | 0.93 | 0.92 | 0.45 | 0.31 | 0.54 | 0.59 | |
| Gemma2-2b LR baselineBackbone=Gemma2-2b, Detection Backend=Logistic Regression baseline, Evaluation Protocol=ID-only training2026.05 | 0.75 | 0.76 | 0.82 | 0.73 | 0.89 | 0.9 | 0.93 | 0.88 | 0.76 | 0.68 | 0.67 | 0.8 | |
| Gemma2-9b LR baselineBackbone=Gemma2-9b, Detection Backend=Logistic Regression baseline, Evaluation Protocol=ID-only training2026.05 | 0.73 | 0.74 | 0.79 | 0.75 | 0.89 | 0.86 | 0.9 | 0.87 | 0.79 | 0.77 | 0.7 | 0.75 |