Multimodal Safety Auditing on RealToxicityPrompts (primary evaluation)
3.05Detoxify Identity Attack ScoreSingle-Prompt Baseline
Evaluation Results
| Method | Links | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Single-Prompt BaselineModality=Multimodal, Prompt Generator=Gemini-1.5-Pro2025.03 | 3.05 | 38.16 | 0 | 21.37 | 1.52 | 55.72 | 55.72 | 4.54 | 37.12 | 1.51 | 15.15 | 4.54 | 48.48 | 55.3 | |
| Greedy-Prompt BaselineModality=Multimodal2025.03 | 3.05 | 39.69 | 0 | 23.66 | 1.52 | 64.12 | 64.12 | 4.54 | 40.9 | 1.51 | 16.66 | 3.78 | 52.27 | 58.33 | |
| RedDiffModality=Multimodal, alignment_weight (lambda)=0, config_label=w/o alignment2025.03 | 3.05 | 45.03 | 0 | 27.48 | 0 | 61.83 | 61.83 | 4.54 | 42.42 | 3.78 | 21.96 | 0.75 | 54.54 | 58.33 | |
| RedDiffModality=Multimodal, alignment_weight (lambda)=0.2, config_label=w/ alignment (H)2025.03 | 3.05 | 40.45 | 0 | 25.19 | 0.76 | 63.35 | 63.35 | 4.54 | 39.39 | 0.75 | 18.93 | 4.54 | 55.3 | 62.12 | |
| RedDiffModality=Multimodal, alignment_weight (lambda)=0.05, config_label=w/ alignment (L)2025.03 | 3.81 | 44.27 | 1.52 | 24.42 | 1.52 | 62.59 | 62.59 | 6.06 | 42.42 | 3.78 | 18.93 | 3.78 | 52.27 | 58.33 | |
| RedDiffModality=Multimodal, alignment_weight (lambda)=0.1, config_label=w/ alignment (M)2025.03 | 3.81 | 42.74 | 0.76 | 25.19 | 0 | 68.7 | 68.7 | 5.3 | 41.66 | 1.51 | 20.45 | 3.03 | 56.81 | 61.36 | |
| Text-Only BaselineModality=Text-Only2025.03 | 5.34 | 39.69 | 0.76 | 19.84 | 0 | 58.01 | 58.01 | 6.06 | 35.6 | 1.51 | 16.66 | 2.27 | 46.96 | 53.03 |