Continual Instruction Tuning on Five Continual Instruction Tuning (CIT) Sequences
68.1Average Accuracy (AA)Oracle Target RL
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Oracle Target RLBackbone=LLaMA-3-8B2026.05 | 68.1 | 7.1 | 70.6 | -4.3 | 87 | |
| PROXYMIXBackbone=LLaMA-3-8B2026.05 | 66.87 | 8.63 | 70.38 | -5.12 | 85.83 | |
| InsCLBackbone=LLaMA-3-8B2026.05 | 63.5 | 12.1 | 68.5 | -8.8 | 80 | |
| Loss-Slope ReplayBackbone=LLaMA-3-8B2026.05 | 63.1 | 12.8 | 68.7 | -9.3 | 79.4 | |
| DoReMi StaticBackbone=LLaMA-3-8B2026.05 | 62.9 | 13.2 | 68.3 | -9.6 | 79.2 | |
| GradSim ReplayBackbone=LLaMA-3-8B2026.05 | 62.7 | 13.4 | 67.9 | -9.8 | 78.6 | |
| Proportional ReplayBackbone=LLaMA-3-8B2026.05 | 62 | 13.9 | 68 | -10.2 | 78.1 | |
| Random ReplayBackbone=LLaMA-3-8B, Replay Ratio=30%2026.05 | 60.1 | 15.9 | 67 | -11.9 | 75.4 | |
| O-LoRABackbone=LLaMA-3-8B2026.05 | 59.4 | 17.5 | 66.3 | -13.1 | 73.6 | |
| Random ReplayBackbone=LLaMA-3-8B, Replay Ratio=10%2026.05 | 57.6 | 20 | 69.8 | -16.1 | 70.2 | |
| No ReplayBackbone=LLaMA-3-8B2026.05 | 48.9 | 32.6 | 72.4 | -29.8 | 54.7 |