Safety Detection on Full (test)
64.3TPR @ 3% FPRGeometry-Lite
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Geometry-LiteBackbone Aggregation=Pooled mean across nine model backbones2026.05 | 64.3 | 75 | |
| MultiLayer-LinearBackbone Aggregation=Pooled mean across nine model backbones2026.05 | 64.1 | 74.7 | |
| TaT-Disp-LSTMBackbone Aggregation=Pooled mean across nine model backbones2026.05 | 56.7 | 69.4 | |
| Best single-layer probeBackbone Aggregation=Pooled mean across nine model backbones2026.05 | 56.4 | 68.6 | |
| MultiLayer-DIMBackbone Aggregation=Pooled mean across nine model backbones2026.05 | 46.6 | 57.4 |