Continual Learning on MLLM-CL (test)
81.1RS ScoreOracle
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Oraclebase_model=LLaVA2025.11 | 81.1 | 65.8 | 54.2 | 56.9 | 91.1 | — | |
| Routingbase_model=LLaVA, approach=Incremental2025.11 | 81 | 66.6 | 53.4 | 56.4 | 91.1 | — | |
| MR-LoRAbase_model=LLaVA, approach=Incremental2025.11 | 80.9 | 65.3 | 54.1 | 56.7 | 91.1 | — | |
| EWCDetails=Elastic Weight Consolidation regularization, LoRA rank=82026.06 | 77.6 | 44.9 | 32.15 | 40.53 | 90.9 | 57.22 | |
| MoELoRAbase_model=LLaVA, approach=Incremental2025.11 | 77.5 | 41.9 | 27.6 | 40.1 | 86.8 | — | |
| SEFEbase_model=LLaVA, approach=Incremental2025.11 | 77.3 | 50.4 | 37.2 | 40.9 | 86.8 | — | |
| Sequential FTDetails=Naive sequential LoRA fine-tuning, LoRA rank=82026.06 | 76.7 | 54.6 | 42.65 | 42.71 | 91.85 | 61.7 | |
| Experience ReplayDetails=Rehearsal with stored past samples, LoRA rank=82026.06 | 76.5 | 54.75 | 43.85 | 43.67 | 91.65 | 62.08 | |
| DISCObase_model=LLaVA, approach=Incremental2025.11 | 76 | 45.2 | 43.8 | 42.3 | 89 | — | |
| MULTIMODAL-REPRODetails=Prototypical Routing, Isolated experts, Signal: Multimodal, LoRA rank=82026.06 | 75.6 | 64.7 | 55.8 | 53.77 | 91.4 | 68.25 | |
| MoELoRADetails=Mixture of LoRA experts with token-level soft routing, LoRA rank=82026.06 | 75.23 | 39.79 | 36.61 | 41.47 | 90.83 | 56.79 | |
| O-LoRAbase_model=LLaVA, approach=Incremental2025.11 | 74.6 | 44.4 | 30 | 41.5 | 87.2 | — | |
| SEFEDetails=Regularized LoRA against semantic forgetting, LoRA rank=82026.06 | 74.38 | 44.6 | 40.24 | 44.41 | 91.47 | 59.02 | |
| HiDebase_model=LLaVA, approach=Incremental2025.11 | 74.3 | 49 | 33.2 | 38.5 | 81.6 | — | |
| MR-LoRADetails=Oracle routing, isolated experts, LoRA rank=82026.06 | 74.2 | 65 | 55.8 | 53.67 | 91.4 | 68.01 | |
| REPRODetails=Prototypical routing, Isolated experts, Signal: Vision, LoRA rank=82026.06 | 73.8 | 64.6 | 55.8 | 53.67 | 91.4 | 67.85 | |
| CL-MoEbase_model=LLaVA, approach=Incremental2025.11 | 71.3 | 46.8 | 26.3 | 41.2 | 88.7 | — | |
| LoRA-FTbase_model=LLaVA, approach=Incremental2025.11 | 69.7 | 41.6 | 25.4 | 40.9 | 87.5 | — | |
| CL-MoEDetails=MoE of LoRAs with dual momentum router, LoRA rank=82026.06 | 69.43 | 34.07 | 31.13 | 38.16 | 86.01 | 51.76 | |
| L2Pbase_model=LLaVA, approach=Incremental2025.11 | 63.8 | 34.6 | 23 | 38.6 | 93 | — | |
| ModalPromptbase_model=LLaVA, approach=Incremental2025.11 | 53.6 | 45.7 | 40.8 | 41.8 | 87.8 | — | |
| Zero-shotbase_model=LLaVA2025.11 | 32.3 | 28.3 | 15.6 | 35.6 | 62.6 | — | |
| Zero-shotDetails=No fine-tuning2026.06 | 32.28 | 28.28 | 15.58 | 43.21 | 62.57 | 36.2 | |
| HiDe-LoRADetails=Hierarchical decomposition of LoRA, LoRA rank=82026.06 | 28.7 | 26 | 18.6 | 36.84 | 52 | 32.43 | |
| DISCO-LoRADetails=Disentangled shared/task-specific LoRA, LoRA rank=82026.06 | 25.5 | 28.75 | 18.7 | 36.04 | 59.65 | 33.73 | |
| O-LoRADetails=Orthogonal subspace LoRA, LoRA rank=82026.06 | 21.97 | 28.2 | 14.19 | 35.48 | 58.67 | 31.7 |