Multimodal Continual Instruction Tuning on TriGap v1 (test)
43.6PMCVQA ScoreCRAM
Evaluation results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| CRAMBackbone=LLaVA-v1.5-7B, Batch size=82026.06 | 43.6 | 44.33 | 16.92 | 66 | 39.17 | 91.83 | 12.83 | 24.27 | 79.49 | 59.53 | 47.79 | |
| DISCOBackbone=LLaVA-v1.5-7B, Batch size=82026.06 | 42.03 | 43.5 | 18.01 | 63.13 | 38.23 | 91.27 | 11.02 | 22.13 | 80.25 | 55.87 | 46.54 | |
| SAMEBackbone=LLaVA-v1.5-7B, Batch size=82026.06 | 41.6 | 43.87 | 17.56 | 64.03 | 39.57 | 90.46 | 10.83 | 21.77 | 81.09 | 54.5 | 46.53 | |
| CL-MoEBackbone=LLaVA-v1.5-7B, Batch size=82026.06 | 40.53 | 36.79 | 13.72 | 52.7 | 32.27 | 93 | 7.77 | 18.33 | 80.09 | 65.9 | 44.11 | |
| MoE-LoRABackbone=LLaVA-v1.5-7B, Batch size=82026.06 | 39.03 | 37.49 | 12.44 | 43.43 | 35.17 | 90.9 | 7.93 | 20.7 | 90.41 | 67 | 44.45 | |
| ModalPromptBackbone=LLaVA-v1.5-7B, Batch size=82026.06 | 38.23 | 38.23 | 11.92 | 44.73 | 37.37 | 84.47 | 10.13 | 12.43 | 71.52 | 52.5 | 40.15 | |
| HiDe-LLaVABackbone=LLaVA-v1.5-7B, Batch size=82026.06 | 37 | 33.2 | 10.52 | 41.97 | 24.09 | 79.2 | 7.73 | 11.17 | 57.39 | 23 | 32.53 | |
| Zero-shotBackbone=LLaVA-v1.5-7B, Zero-shot=true2026.06 | 35.4 | 12.68 | 9.36 | 19.27 | 5.06 | 53.77 | 7.4 | 5.3 | 47.41 | 20.37 | 21.6 | |
| FT-LoRABackbone=LLaVA-v1.5-7B, Batch size=82026.06 | 34.2 | 23.32 | 9.84 | 37.07 | 23.53 | 83.83 | 7 | 12.7 | 80.31 | 60.27 | 37.21 | |
| Replay-LoRABackbone=LLaVA-v1.5-7B, Batch size=82026.06 | 33.7 | 33.95 | 14 | 46.67 | 28.97 | 75.57 | 9.4 | 15.9 | 73.81 | 58.8 | 39.08 |