Mathematical Reasoning on GSM8K (Accuracy, Attention Recall)
81Accuracy (GSM8K)Full
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| FullBackbone=LLaMA-3.1-8B-Instruct2026.06 | 81 | 100 | |
| RaBitQCacheBackbone=LLaMA-3.1-8B-Instruct, Top-p=0.92026.06 | 77 | 88.8 | |
| SparQBackbone=LLaMA-3.1-8B-Instruct, Configuration=default setting2026.06 | 77 | 60.5 | |
| QuestBackbone=LLaMA-3.1-8B-Instruct, Token Budget=2562026.06 | 70 | 81.3 |