Knowledge Preservation on Knowledge Datasets (ARC-C, ARC-E, MMLU, GPQA)
64.75ARC-CReasoning
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| ReasoningBackbone=Qwen2.5-7B2026.01 | 64.75 | 77.25 | 52.51 | 49.1 | 64.89 | |
| ReasonAnyBackbone=Qwen2.5-7B2026.01 | 64.31 | 73.39 | 72.73 | 43.18 | 59.33 | |
| Task ArithmeticBackbone=Qwen2.5-7B2026.01 | 63.56 | 70.43 | 62.14 | 35.61 | 44.63 | |
| SafetyBackbone=Qwen2.5-7B2026.01 | 60.91 | 65.78 | 71.72 | 39.39 | 50.85 | |
| LinearBackbone=Qwen2.5-7B2026.01 | 60.34 | 66.19 | 57.34 | 28.78 | 37.82 | |
| LEDBackbone=Qwen2.5-7B2026.01 | 32.54 | 33.69 | 71.93 | 38.64 | 44.14 | |
| TiesBackbone=Qwen2.5-7B2026.01 | 21.36 | 26.63 | 23.08 | 29.55 | 13.46 | |
| DAREBackbone=Qwen2.5-7B2026.01 | 17.97 | 13.93 | 25.95 | 3.03 | 7.31 | |
| FuseLLMBackbone=Qwen2.5-7B2026.01 | 0 | 0 | 22.95 | 0 | 3.4 |