Mathematical Reasoning on GSM8K (Accuracy, Performance Gain)
96.81Accuracyl2-guided Selection
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| l2-guided SelectionModel=Qwen3-32B2026.06 | 96.81 | 0.38 | |
| Base (Mean@8)Model=Qwen3-32B2026.06 | 96.44 | — | |
| l2-guided SelectionModel=Qwen3-14B2026.06 | 96.36 | 0.24 | |
| l2-guided SelectionModel=Qwen3-4B2026.06 | 96.2 | 0.83 | |
| Base (Mean@8)Model=Qwen3-14B2026.06 | 96.13 | — | |
| l2-guided SelectionModel=Qwen3-8B2026.06 | 96.06 | 0.16 | |
| Base (Mean@8)Model=Qwen3-8B2026.06 | 95.91 | — | |
| Base (Mean@8)Model=Qwen3-4B2026.06 | 95.41 | — | |
| l2-guided SelectionModel=LLaMA-70B2026.06 | 95.14 | 0.88 | |
| Base (Mean@8)Model=LLaMA-70B2026.06 | 94.31 | — | |
| l2-guided SelectionModel=Qwen3-1.7B2026.06 | 91.21 | 1.31 | |
| l2-guided SelectionModel=LLaMA-8B2026.06 | 90.82 | 2.81 | |
| Base (Mean@8)Model=Qwen3-1.7B2026.06 | 90.03 | — | |
| Base (Mean@8)Model=LLaMA-8B2026.06 | 88.34 | — |