Safety-shift experiment on RealToxicityPrompts
90CoverageAdaptive CRC
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Adaptive CRC2026.06 | 90 | 500 | 0 | |
| PromptShift-CRC2026.06 | 90 | 618 | 0 | |
| Representation-weighted CRC2026.06 | 89.7 | 548.7 | 0.3 | |
| Rolling CRC2026.06 | 89.6 | 500 | 0.4 | |
| Static CRC2026.06 | 89.1 | 600 | 0.9 |