Multimodal Knowledge Retention on Retain-Set 15% forgetting ratio experiment
61.6Classification VQA AccuracyBefore
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| BeforeBackbone=Qwen3-VL-8B, Forgetting Ratio=15%2026.05 | 61.6 | 61.78 | 71.3 | 75.5 | 20.5 | 29.7 | |
| MMunlearnerBackbone=Qwen3-VL-8B, Forgetting Ratio=15%2026.05 | 61.33 | 63.63 | 75.9 | 76.8 | 21.5 | 31 | |
| KL_MinBackbone=Qwen3-VL-8B, Forgetting Ratio=15%2026.05 | 61.22 | 61.05 | 68.6 | 74.3 | 19.9 | 29.4 | |
| GABackbone=Qwen3-VL-8B, Forgetting Ratio=15%2026.05 | 59.58 | 60.58 | 26.6 | 70 | 17.4 | 27.8 | |
| GA_diffBackbone=Qwen3-VL-8B, Forgetting Ratio=15%2026.05 | 59.3 | 61.94 | 75.6 | 75.7 | 21.3 | 30.1 | |
| MANUBackbone=Qwen3-VL-8B, Forgetting Ratio=15%2026.05 | 53.77 | 61.58 | 61.3 | 73.6 | 15 | 24.8 | |
| NPOBackbone=Qwen3-VL-8B, Forgetting Ratio=15%2026.05 | 52.81 | 56.31 | 26.6 | 60.5 | 15.5 | 25.3 |