Continual Multimodal Learning on MRCL Benchmark (test)
85.57MedBook ScoreSEFE
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| SEFEBackbone=Qwen3-VL-2B2026.07 | 85.57 | 20.59 | 35.63 | 81.25 | 34.2 | 63.62 | 51.45 | 57.57 | |
| GRPO*Backbone=Qwen3-VL-2B2026.07 | 76.23 | 17.1 | 51.24 | 61.5 | 42.15 | 53.72 | 49.64 | 51.59 | |
| CPOBackbone=Qwen3-VL-2B2026.07 | 76.06 | 43.38 | 49.03 | 78 | 42.65 | 68.28 | 57.82 | 62.28 | |
| O-LoRABackbone=Qwen3-VL-2B2026.07 | 69.61 | 12.5 | 33.49 | 85 | 34.4 | 61.75 | 47 | 55.42 | |
| GSPOBackbone=Qwen3-VL-2B2026.07 | 69.27 | 11.95 | 45.2 | 75.75 | 41.05 | 68.23 | 48.64 | 57.96 | |
| KeepLoRABackbone=Qwen3-VL-2B2026.07 | 63.5 | 35.29 | 33.95 | 86 | 31.05 | 63.84 | 49.96 | 57.21 | |
| GRPOBackbone=Qwen3-VL-2B2026.07 | 62.48 | 28.86 | 48.11 | 66.25 | 41.1 | 60.86 | 49.36 | 55.53 | |
| LoRABackbone=Qwen3-VL-2B2026.07 | 52.8 | 18.01 | 35.47 | 85 | 32.35 | 63.45 | 44.73 | 53.97 | |
| Zero-shot¹Backbone=Qwen3-VL-2B, Protocol=Zero-shot2026.07 | 52.63 | 0 | 17.23 | 35.75 | 32.75 | — | — | 27.67 | |
| Zero-shot²Backbone=Qwen3-VL-2B, Protocol=Zero-shot2026.07 | 47.2 | 1.65 | 40.4 | 22 | 31.45 | — | — | 28.54 | |
| Full SFTBackbone=Qwen3-VL-2B2026.07 | 44.65 | 0.55 | 33.41 | 83.25 | 33.6 | 66.32 | 39.09 | 52.02 |