Forgetting Analysis on Multi-Dataset Suite (GSM8K, Code, LogiQA, Alpaca, UChat)
4.8GSM8K Forgetting RateEPI
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| EPIBase Model=Gemma-2-9B, p=1%2026.04 | 4.8 | 4.5 | 4.2 | 3.8 | 3.4 | 4.14 | |
| EPIBase Model=LLaMA-3-8B, p=1%2026.04 | 5.3 | 5.1 | 4.8 | 4.2 | 3.9 | 4.66 | |
| Static IsolationBase Model=Gemma-2-9B, p=1%2026.04 | 5.6 | 5.3 | 4.9 | 4.1 | 3.8 | 4.74 | |
| EPIBase Model=Qwen2-7B, p=1%2026.04 | 5.7 | 5 | 4.7 | 4.1 | 3.8 | 4.66 | |
| Static IsolationBase Model=LLaMA-3-8B, p=1%2026.04 | 6.1 | 5.8 | 5.2 | 4.5 | 4.1 | 5.14 | |
| Static IsolationBase Model=Qwen2-7B, p=1%2026.04 | 6.8 | 6.3 | 5.8 | 4.9 | 4.5 | 5.66 | |
| Multi-Stage (Heuristic)Base Model=Gemma-2-9B2026.04 | 7.9 | 7.2 | 6.8 | 6.2 | 5.7 | 6.36 | |
| Multi-Stage (Random)Base Model=Gemma-2-9B2026.04 | 8.5 | 7.9 | 7.3 | 6.5 | 6 | 7.04 | |
| Multi-Stage (Heuristic)Base Model=LLaMA-3-8B2026.04 | 8.9 | 8.2 | 7.5 | 6.5 | 6 | 7.42 | |
| Multi-Stage (Heuristic)Base Model=Qwen2-7B2026.04 | 9.4 | 8.5 | 8 | 6.7 | 6.1 | 7.74 | |
| Multi-Stage (Random)Base Model=LLaMA-3-8B2026.04 | 9.7 | 8.9 | 8.1 | 6.8 | 6.2 | 7.94 | |
| Multi-Stage (Random)Base Model=Qwen2-7B2026.04 | 10.1 | 9.1 | 8.5 | 6.9 | 6.4 | 8.2 | |
| Full SFT (Multi-task)Base Model=Gemma-2-9B2026.04 | 11.2 | 9.7 | 9 | 7.3 | 7.1 | 8.66 | |
| Full SFT (Multi-task)Base Model=LLaMA-3-8B2026.04 | 12.6 | 10.4 | 9.8 | 7.5 | 6.9 | 9.44 | |
| Full SFT (Multi-task)Base Model=Qwen2-7B2026.04 | 13 | 11.5 | 10.2 | 8 | 7.5 | 10.04 |