Multimodal Machine Unlearning on CLEAR
32.4Forget VQA AccuracyCVF
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| CVFBackbone=LLaVA-1.5-7B2026.05 | 32.4 | 0.367 | 50 | 0.352 | 53 | 85.4 | |
| MMUBackbone=LLaVA-1.5-7B2026.05 | 36.2 | 0.348 | 46.6 | 0.338 | 52.3 | 84.1 | |
| MANUBackbone=LLaVA-1.5-7B2026.05 | 39.8 | 0.263 | 25.8 | 0.224 | 49.8 | 68.3 | |
| KL_MinBackbone=LLaVA-1.5-7B2026.05 | 40.4 | 0.27 | 38.1 | 0.274 | 51.5 | 82.8 | |
| NPOBackbone=LLaVA-1.5-7B2026.05 | 40.4 | 0.285 | 38.6 | 0.282 | 52.9 | 83.4 | |
| GA_DiffBackbone=LLaVA-1.5-7B2026.05 | 47.3 | 0.197 | 43.4 | 0.22 | 47.7 | 73.5 | |
| CVFBackbone=Qwen2-VL-7B2026.05 | 47.6 | 0.116 | 70.9 | 0.098 | 69 | 91.4 | |
| MMUBackbone=Qwen2-VL-7B2026.05 | 50 | 0.123 | 68.9 | 0.094 | 68.9 | 87.3 | |
| GABackbone=Qwen2-VL-7B2026.05 | 55.3 | 0.123 | 62.4 | 0.083 | 65.9 | 86.8 | |
| GABackbone=LLaVA-1.5-7B2026.05 | 57.4 | 0.153 | 48.4 | 0.176 | 51.8 | 83.4 | |
| MANUBackbone=Qwen2-VL-7B2026.05 | 61.3 | 0.107 | 65.2 | 0.094 | 66.2 | 87.3 | |
| NPOBackbone=Qwen2-VL-7B2026.05 | 62.8 | 0.103 | 68.3 | 0.091 | 68.9 | 88.7 | |
| VanillaBackbone=LLaVA-1.5-7B2026.05 | 63.3 | 0.367 | 54 | 0.352 | 53.7 | 85.4 | |
| GA_DiffBackbone=Qwen2-VL-7B2026.05 | 63.3 | 0.125 | 67.4 | 0.088 | 68.4 | 92.7 | |
| VanillaBackbone=Qwen2-VL-7B2026.05 | 67 | 0.116 | 70.9 | 0.098 | 69.2 | 91.4 | |
| KL_MinBackbone=Qwen2-VL-7B2026.05 | 67 | 0.12 | 68.9 | 0.098 | 68.4 | 90.7 |