Mathematical Reasoning on GSM8K (Accuracy & Max Accuracy)
58.4AccuracyReSCALE
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ReSCALEBudget=Large, Tokens=16K–18K, Backbone=Llama2-7B2026.03 | 58.4 | 59.7 | |
| ReSCALEBudget=Medium, Tokens=6K–8K, Backbone=Llama2-7B2026.03 | 57.9 | 60.2 | |
| AlphaZeroBudget=Medium, Tokens=6K–8K, Backbone=Llama2-7B2026.03 | 56.7 | 58.3 | |
| AlphaZeroBudget=Small, Tokens=0.5K–2K, Backbone=Llama2-7B2026.03 | 56 | 58.8 | |
| ReSCALEBudget=Small, Tokens=0.5K–2K, Backbone=Llama2-7B2026.03 | 55.1 | 58.6 | |
| AlphaZeroBudget=Large, Tokens=16K–18K, Backbone=Llama2-7B2026.03 | 53.6 | 55 | |
| Best-of-NBudget=N = 32, Tokens=3.5K, Backbone=Llama2-7B2026.03 | 53.4 | — |