Attack Success Rate Correlation Analysis on GCG/AutoDAN sparse-transfer diagnostic
0.965Absolute RhoTarget HRCI
Evaluation Results
| Method | Links | |
|---|---|---|
| Target HRCIRegime=R2D2, Family=geometry2026.06 | 0.965 | |
| Target fixed-source ASRRegime=R2D2, Family=behavior state2026.06 | 0.965 | |
| Target-source HRCI gapRegime=R2D2, Family=geometry gap2026.06 | 0.735 | |
| Target refusal-score driftRegime=SFT, Family=drift only2026.06 | 0.591 | |
| Target fixed-source ASRRegime=SFT, Family=behavior state2026.06 | 0.562 | |
| Target refusal-score driftRegime=R2D2, Family=drift only2026.06 | 0.482 | |
| Target-source refusal-drift gapRegime=SFT, Family=drift gap2026.06 | 0.424 | |
| Target HRCIRegime=SFT, Family=geometry2026.06 | 0.03 |