Adversarial Detection on Wikipedia HiSPA word-shuffle benign adversarial (N=500 held-out)
65F1 ScoreSpectralGuard
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| SpectralGuardModel=Mamba-2.8B, Params=2,897M2026.03 | 65 | 72.4 | 0.8787 | 0.8778 | 0.0009 | |
| SpectralGuardModel=Mamba-130M, Params=130M2026.03 | 61.9 | 67.2 | 0.9086 | 0.909 | 0.0004 | |
| SpectralGuardModel=Mamba-1.4B, Params=1,475M2026.03 | 59.1 | 67.2 | 0.9205 | 0.9211 | 0.0006 |