Prompt Selection on Perturbation-100 (test)
76Top-1 HitRisk-Cos
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Risk-CosMode=Prefill-only2026.03 | 76 | 18.8 | 0.604 | 0.55 | 20.8 | 0.17 | |
| Risk-MarginMode=Prefill-only2026.03 | 76 | 32.5 | 0.481 | 0.43 | -3.8 | -0.02 | |
| Risk-EntropyMode=Prefill-only2026.03 | 57 | 65.7 | 0.381 | 0.33 | -23.9 | -0.18 | |
| Risk-Loss (NLL)Mode=Prefill-only2026.03 | 52 | 87 | 0.359 | 0.31 | -28.2 | -0.23 |