Mathematical Reasoning on GSM8K (Accuracy, Time)
48.3AccuracyFor-Value
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| For-ValueBackbone=Llama-3.1-8B, Selection Ratio=5%2025.08 | 48.3 | 0.3 | |
| Full (100%)Backbone=Llama-3.1-8B, Selection Ratio=100%2025.08 | 47.8 | — | |
| For-ValueBackbone=Llama-3.1-8B, Selection Ratio=1%2025.08 | 45.2 | 0.3 | |
| HyperINFBackbone=Llama-3.1-8B, Selection Ratio=5%2025.08 | 42.8 | 2.4 | |
| DataInfBackbone=Llama-3.1-8B, Selection Ratio=5%2025.08 | 42 | 1.9 | |
| HyperINFBackbone=Llama-3.1-8B, Selection Ratio=1%2025.08 | 41.9 | 2.4 | |
| Hessian-freeBackbone=Llama-3.1-8B, Selection Ratio=5%2025.08 | 41.8 | 1.4 | |
| DataInfBackbone=Llama-3.1-8B, Selection Ratio=1%2025.08 | 41.7 | 1.9 | |
| Hessian-freeBackbone=Llama-3.1-8B, Selection Ratio=1%2025.08 | 41.5 | 1.4 |