Multimodal Continual Learning on MLLM-CTBENCH
55.17Accuracy (Math QA)SeqLoRA
Evaluation Results
| Method | Links | ||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SeqLoRA2026.01 | 55.17 | — | 11.62 | -25.76 | 29.89 | -23.93 | 31.97 | -40.51 | 17.54 | -19.18 | 9.19 | -15.74 | 43.7 | -42.63 | 28.44 | -23.96 | |
| O-LoRA2026.01 | 54.84 | — | 29.16 | -5 | 35.48 | -16.63 | 34.53 | -37.45 | 23.63 | -10.16 | 13.49 | -8.66 | 56.84 | -25.35 | 35.42 | -14.75 | |
| SeqFT2026.01 | 53.94 | — | 9.08 | -24.99 | 36.57 | -14.87 | 35.65 | -37.48 | 20.01 | -16.48 | 11.47 | -14.41 | 56.84 | -28.63 | 33.93 | -17.56 | |
| EWC2026.01 | 53.94 | — | 2.27 | -31.25 | 30.56 | -22.12 | 37.9 | -35.29 | 22.93 | -12.09 | 2.27 | -20.17 | 50.99 | -31.2 | 28.69 | -21.74 | |
| PASs-MoE2026.01 | 53.69 | — | 60.25 | 26.99 | 46.52 | -4.34 | 66.94 | -1.61 | 28.11 | -6.43 | 6.25 | -16.04 | 64.94 | -17.72 | 46.67 | -2.74 | |
| MoELoRAvariant=Softmax2026.01 | 53.45 | — | 2.49 | -31.05 | 31.24 | -20.19 | 38.71 | -35.08 | 22.94 | -11.81 | 2.27 | -20.38 | 50.42 | -32.14 | 28.79 | -21.52 | |
| MAS2026.01 | 53.2 | — | 2.63 | -30.92 | 30.22 | -21.32 | 37.1 | -36.49 | 22.7 | -12.5 | 2.84 | -19.6 | 51.08 | -31.48 | 28.54 | -21.76 | |
| Replay2026.01 | 52.71 | — | 23.41 | -10.5 | 35.92 | -11.4 | 61.29 | -11.49 | 27.81 | -5.87 | 5.68 | -16.74 | 53.63 | -28.93 | 37.21 | -12.13 | |
| MoELoRAvariant=Top-k2026.01 | 51.56 | — | 2.47 | -30.61 | 29.66 | -19.08 | 34.27 | -32.92 | 16.08 | -16.39 | 3.47 | -18.32 | 42.86 | -35.62 | 25.77 | -21.85 | |
| LwF2026.01 | 45.1 | — | 5.21 | -28.42 | 36.36 | -15.07 | 32.16 | -39.01 | 21.77 | -11.25 | 19.84 | -1.95 | 52.4 | -30.35 | 30.41 | -18 | |
| HiDe-llava2026.01 | 42.12 | — | 28.33 | -2.74 | 43.33 | -1.82 | 47.78 | -14.82 | 24.87 | -7.7 | 15.17 | -8.62 | 46.65 | -36.57 | 28.8 | -16.99 | |
| DDAS2026.01 | 40.89 | — | 13.48 | -0.48 | 37.06 | 10.38 | 54.64 | -13.71 | 21.65 | -2.82 | 1.07 | -0.78 | 50.05 | -13.95 | 31.26 | -3.05 |