Safety Detection on WildJailbreak (held-out)
99AUROCGeometry-Lite
Evaluation Results
| Method | Links | |
|---|---|---|
| Geometry-LiteProtocol=Leave-one-benchmark-out (LOBO), Aggregation=Mean across nine model backbones2026.05 | 99 | |
| TaT-Disp-LSTMProtocol=Leave-one-benchmark-out (LOBO), Aggregation=Mean across nine model backbones2026.05 | 98.4 | |
| MultiLayer-LinearProtocol=Leave-one-benchmark-out (LOBO), Aggregation=Mean across nine model backbones2026.05 | 96.5 | |
| MultiLayer-DIMProtocol=Leave-one-benchmark-out (LOBO), Aggregation=Mean across nine model backbones2026.05 | 92 | |
| Best single-layer probeProtocol=Leave-one-benchmark-out (LOBO), Aggregation=Mean across nine model backbones2026.05 | 83.5 |