Multimodal Machine Unlearning on Forget Set (5%)
44.4Classification AccuracyVanilla
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| VanillaBase Model=LLaVA-13B, Forget Ratio=5%2025.02 | 44.4 | 0.604 | 6.1 | 21.9 | — | — | — | — | — | — | |
| PromptingBase Model=LLaVA-13B, Forget Ratio=5%2025.02 | 42.99 | 0.588 | 5.36 | 20.8 | — | — | — | — | — | — | |
| KL Minim.Base Model=LLaVA-13B, Forget Ratio=5%2025.02 | 38.64 | 0.561 | 4.17 | 19.97 | — | — | — | — | — | — | |
| NPOBase Model=LLaVA-13B, Forget Ratio=5%2025.02 | 35.12 | 0.473 | 3.41 | 18.77 | — | — | — | — | — | — | |
| Grad. DiffBase Model=LLaVA-13B, Forget Ratio=5%2025.02 | 34.16 | 0.435 | 3.95 | 15.55 | — | — | — | — | — | — | |
| GABase Model=LLaVA-13B, Forget Ratio=5%2025.02 | 34.04 | 0.423 | 2.98 | 15.19 | — | — | — | — | — | — | |
| MANUBase Model=LLaVA-13B, Forget Ratio=5%2025.02 | 33.65 | 0.434 | 3.13 | 15.33 | — | — | — | — | — | — | |
| BeforeBackbone=Qwen3-VL-8B, Forgetting Ratio=5%2026.05 | — | — | — | — | 62.37 | 63.8 | 0.687 | 0.733 | 19.8 | 31.6 | |
| GABackbone=Qwen3-VL-8B, Forgetting Ratio=5%2026.05 | — | — | — | — | 61.2 | 62.54 | 0.672 | 0.736 | 21.1 | 30.7 | |
| GA_diffBackbone=Qwen3-VL-8B, Forgetting Ratio=5%2026.05 | — | — | — | — | 46.24 | 64.16 | 0.478 | 0.725 | 15.9 | 30.3 | |
| KL_MinBackbone=Qwen3-VL-8B, Forgetting Ratio=5%2026.05 | — | — | — | — | 62.54 | 63.26 | 0.688 | 0.732 | 19.9 | 32.1 | |
| MANUBackbone=Qwen3-VL-8B, Forgetting Ratio=5%2026.05 | — | — | — | — | 49.64 | 64.87 | 0.572 | 0.7061 | 13.77 | 23.7 | |
| MMunlearnerBackbone=Qwen3-VL-8B, Forgetting Ratio=5%2026.05 | — | — | — | — | 55.73 | 59.14 | 0.633 | 0.719 | 19.1 | 27.9 | |
| NPOBackbone=Qwen3-VL-8B, Forgetting Ratio=5%2026.05 | — | — | — | — | 40.86 | 54.48 | 0.196 | 0.373 | 7.9 | 13.6 |