Multimodal Machine Unlearning on MLLMU-Bench
29.6Forget VQA AccuracyCVF
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| CVFBackbone=LLaVA-1.5-7B2026.05 | 29.6 | 38.4 | 45 | 37.5 | 47 | 54.9 | |
| MMUBackbone=LLaVA-1.5-7B2026.05 | 31.2 | 34.2 | 44.2 | 35.1 | 46.7 | 54.9 | |
| GA_DiffBackbone=LLaVA-1.5-7B2026.05 | 40 | 33.6 | 44.3 | 31.5 | 46.6 | 53.6 | |
| MANUBackbone=LLaVA-1.5-7B2026.05 | 41.6 | 33.6 | 43.2 | 32.1 | 46.3 | 54.8 | |
| KL_MinBackbone=LLaVA-1.5-7B2026.05 | 42.4 | 33.6 | 44.9 | 32 | 47.4 | 54.6 | |
| CVFBackbone=Qwen2-VL-7B2026.05 | 42.8 | 55 | 56 | 58.6 | 75.4 | 77.5 | |
| GABackbone=LLaVA-1.5-7B2026.05 | 43.2 | 32.5 | 45 | 32.2 | 47 | 54.7 | |
| NPOBackbone=LLaVA-1.5-7B2026.05 | 43.2 | 33.6 | 44.9 | 32.2 | 47 | 54.7 | |
| MANUBackbone=Qwen2-VL-7B2026.05 | 43.2 | 47.2 | 43.2 | 49.4 | 73.7 | 76.2 | |
| MMUBackbone=Qwen2-VL-7B2026.05 | 44 | 54.4 | 47.4 | 55.7 | 75.2 | 77.3 | |
| KL_MinBackbone=Qwen2-VL-7B2026.05 | 45.6 | 45.3 | 35.9 | 55.6 | 74.8 | 77.1 | |
| VanillaBackbone=LLaVA-1.5-7B2026.05 | 45.8 | 38.4 | 45.2 | 37.5 | 47.4 | 54.9 | |
| NPOBackbone=Qwen2-VL-7B2026.05 | 49.6 | 50.4 | 49.5 | 53.3 | 75.2 | 78.3 | |
| GABackbone=Qwen2-VL-7B2026.05 | 50.4 | 46.7 | 51.5 | 57.6 | 74.4 | 77.8 | |
| GA_DiffBackbone=Qwen2-VL-7B2026.05 | 54.4 | 52.8 | 38.8 | 54.4 | 74.5 | 77 | |
| VanillaBackbone=Qwen2-VL-7B2026.05 | 55.2 | 55 | 56 | 58.6 | 77.3 | 77.5 |