Continual Learning on MTCL-Bench
55.97AccuracyPer-task LoRA SFT
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Per-task LoRA SFTModel=Qwen3-4B Instruct, Is CL?=false2025.07 | 55.97 | — | |
| ProgPromptModel=Qwen3-4B Instruct, Is CL?=true2025.07 | 54.85 | -6.12 | |
| GRIDModel=Qwen3-4B Instruct, Is CL?=true2025.07 | 53.63 | -5.14 | |
| Per-task LoRA SFTModel=LLaMA-3.1 8B-Instruct, Is CL?=false2025.07 | 51.53 | — | |
| GRIDModel=LLaMA-3.1 8B-Instruct, Is CL?=true2025.07 | 50.55 | -25.34 | |
| ProgPromptModel=LLaMA-3.1 8B-Instruct, Is CL?=true2025.07 | 49.15 | -35.09 | |
| PretrainedModel=LLaMA-3.1 8B-Instruct, Is CL?=false2025.07 | 45.9 | — | |
| PretrainedModel=Qwen3-4B Instruct, Is CL?=false2025.07 | 45.16 | — | |
| Prompt TuningModel=LLaMA-3.1 8B-Instruct, Is CL?=true2025.07 | 16.89 | -42.67 | |
| Prompt TuningModel=Qwen3-4B Instruct, Is CL?=true2025.07 | 14.12 | -7.34 |