Conformal Risk Control on Synthetic prompt-shift stream
89.2CoveragePromptShift-CRC
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| PromptShift-CRCTarget coverage=0.902026.06 | 89.2 | 0.757 | 2.621 | 232.3 | |
| Adaptive CRCTarget coverage=0.902026.06 | 88.2 | 0.761 | 2.383 | 300 | |
| Rolling CRCTarget coverage=0.902026.06 | 85.1 | 0.768 | 2.383 | 300 | |
| Representation-weighted CRCTarget coverage=0.902026.06 | 82.8 | 0.735 | 2.617 | 297.6 | |
| Static CRCTarget coverage=0.902026.06 | 34.4 | 0.505 | 2.641 | 300 |