Multi-modal Toxicity Attack on RealToxicityPrompts (RTP) (test)
31.36Overall ScoreSTARE (0.2)
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| STARE (0.2)Access Tier (Tier-1 vs Tier-2)=Tier-1, Alignment weight=0.22026.05 | 31.36 | 17.1 | 6.29 | 29.73 | 4.38 | 15.95 | 6.14 | 0.78 | |
| STARE (0.05)Access Tier (Tier-1 vs Tier-2)=Tier-1, Alignment weight=0.052026.05 | 28.93 | 16.84 | 5.72 | 26.57 | 3.9 | 16.69 | 5.77 | 0.72 | |
| STARE (0.1)Access Tier (Tier-1 vs Tier-2)=Tier-1, Alignment weight=0.12026.05 | 28.54 | 17.08 | 5.77 | 25.75 | 4.22 | 16.14 | 6.82 | 0.76 | |
| STARE w/ DDPOAccess Tier (Tier-1 vs Tier-2)=Tier-1, Access assumption (cross-paradigm vs controlled)=controlled; identical white-box T2I budget as STARE, Configuration=with DDPO2026.05 | 27.84 | 15.62 | 5.41 | 26.12 | 3.92 | 15.11 | 5.8 | 0.75 | |
| STARE w/o AlignAccess Tier (Tier-1 vs Tier-2)=Tier-1, Component removal (Ablation)=w/o Align2026.05 | 26.43 | 14.84 | 5.09 | 25.8 | 3.95 | 14.05 | 5.83 | 0.68 | |
| STARE w/o EditAccess Tier (Tier-1 vs Tier-2)=Tier-1, Component removal (Ablation)=w/o Edit2026.05 | 25.56 | 13.46 | 4.68 | 24.88 | 3.27 | 10.56 | 5.5 | 0.72 | |
| STARE w/o LoRAAccess Tier (Tier-1 vs Tier-2)=Tier-1, Component removal (Ablation)=w/o LoRA2026.05 | 22.04 | 12.13 | 5.11 | 19.77 | 4.09 | 11.06 | 5.99 | 0.75 | |
| ARTAccess Tier (Tier-1 vs Tier-2)=Tier-2, Access assumption (cross-paradigm vs controlled)=cross-paradigm; different access, for context2026.05 | 18.62 | 9.22 | 2.15 | 17.54 | 4.05 | 8.94 | 6.45 | 0.75 | |
| DiffZOOAccess Tier (Tier-1 vs Tier-2)=Tier-2, Access assumption (cross-paradigm vs controlled)=cross-paradigm; different access, for context2026.05 | 17.2 | 9.01 | 2.12 | 16.42 | 2.72 | 7.88 | 4.14 | 0.73 | |
| PGJAccess Tier (Tier-1 vs Tier-2)=Tier-2, Access assumption (cross-paradigm vs controlled)=cross-paradigm; different access, for context2026.05 | 14.86 | 7.85 | 1.75 | 13.98 | 1.65 | 8.09 | 3.43 | 0.71 | |
| Text + SDAccess Tier (Tier-1 vs Tier-2)=Tier-2, Access assumption (cross-paradigm vs controlled)=cross-paradigm; different access, for context2026.05 | 11.15 | 5.71 | 0.38 | 10.63 | 0.22 | 6.11 | 3.97 | 0.72 | |
| Text-OnlyAccess Tier (Tier-1 vs Tier-2)=Tier-2, Access assumption (cross-paradigm vs controlled)=cross-paradigm; different access, for context2026.05 | 5.2 | 3.1 | 0.2 | 5.1 | 0.8 | 2.8 | 0.6 | — |