Visual Question Answering Unlearning on MLLMU-Bench 10% forget ratio (test)
56.67Forget Quality (Class - Forget)MMunlearner
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| MMunlearnersamples (Dr)=100%, Backbone=Qwen3-VL-8B-Instruct2026.05 | 56.67 | 60.83 | 54 | 31.5 | 0.3 | 0.34 | 52.96 | 73.11 | 52.6 | 37.6 | |
| NPOsamples (Dr)=0.00%, Backbone=Qwen3-VL-8B-Instruct2026.05 | 47.75 | 54.69 | 46.1 | 33.1 | 0.01 | 0.04 | 51.07 | 72.54 | 46 | 38.5 | |
| GAsamples (Dr)=0.00%, Backbone=Qwen3-VL-8B-Instruct2026.05 | 46.12 | 53.06 | 56.9 | 36.9 | 0.01 | 0.01 | 53.08 | 71.54 | 55.6 | 38.7 | |
| GA_diffsamples (Dr)=100%, Backbone=Qwen3-VL-8B-Instruct2026.05 | 46.12 | 51.43 | 52.3 | 33 | 0.05 | 0.05 | 49.78 | 72.06 | 54.1 | 38.6 | |
| Vanillasamples (Dr)=-, Backbone=Qwen3-VL-8B-Instruct2026.05 | 45.71 | 54.29 | 56.9 | 35.5 | — | — | 52.99 | 78.59 | 57 | 43.4 | |
| KL_Minsamples (Dr)=100%, Backbone=Qwen3-VL-8B-Instruct2026.05 | 43.27 | 53.06 | 57.4 | 32.6 | 0.07 | 0.09 | 49.95 | 71.48 | 55.2 | 37.4 | |
| ASRUsamples (Dr)=28.80%, Backbone=Qwen3-VL-8B-Instruct2026.05 | 35.92 | 48.97 | 27.1 | 25.7 | 4.54 | 4.41 | 53.93 | 73.24 | 55.9 | 38.9 |