Visual Question Answering on MLLMU-Bench 5% forget (test)
62.5Classification AccuracyVanilla
Evaluation results
| Method | Links | ||
|---|---|---|---|
| VanillaBackbone=Qwen3-VL-8B, Samples=-2026.05 | 62.5 | 37 | |
| NPOBackbone=Qwen3-VL-8B, Samples=0%2026.05 | 56.67 | 31.4 | |
| GABackbone=Qwen3-VL-8B, Samples=0%2026.05 | 55 | 33 | |
| GA_diffBackbone=Qwen3-VL-8B, Samples=100%2026.05 | 55 | 33.4 | |
| KL_MinBackbone=Qwen3-VL-8B, Samples=100%2026.05 | 53.33 | 35.5 | |
| MMunlearnerBackbone=Qwen3-VL-8B, Samples=100%2026.05 | 52 | 34.6 | |
| VanillaBackbone=Qwen3-VL-4B, Samples=-2026.05 | 50 | 32.7 | |
| GA_diffBackbone=Qwen3-VL-4B, Samples=100%2026.05 | 50 | 31.2 | |
| MMunlearnerBackbone=Qwen3-VL-4B, Samples=100%2026.05 | 47.2 | 29.8 | |
| KL_MinBackbone=Qwen3-VL-4B, Samples=100%2026.05 | 46.67 | 31.5 | |
| GABackbone=Qwen3-VL-4B, Samples=0%2026.05 | 45 | 28.8 | |
| ASRUBackbone=Qwen3-VL-8B, Samples=27.30%2026.05 | 45 | 27.7 | |
| NPOBackbone=Qwen3-VL-4B, Samples=0%2026.05 | 43.33 | 32.7 | |
| MMunlearnerBackbone=LLaVA-1.5-7B, Samples=100%2026.05 | 42.5 | 18.1 | |
| VanillaBackbone=LLaVA-1.5-7B, Samples=-2026.05 | 41.67 | 23 | |
| GABackbone=LLaVA-1.5-7B, Samples=0%2026.05 | 41.67 | 23 | |
| GA_diffBackbone=LLaVA-1.5-7B, Samples=100%2026.05 | 40.83 | 17 | |
| KL_MinBackbone=LLaVA-1.5-7B, Samples=100%2026.05 | 40.83 | 22.7 | |
| NPOBackbone=LLaVA-1.5-7B, Samples=0%2026.05 | 40 | 21.5 | |
| ASRUBackbone=LLaVA-1.5-7B, Samples=27.30%2026.05 | 38.4 | 14.4 | |
| ASRUBackbone=Qwen3-VL-4B, Samples=27.30%2026.05 | 34.4 | 24.5 |