Arithmetic Reasoning on Game of 24
85.3PerformanceReSCALE
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ReSCALEBudget=Large, Tokens=4K–6K, Backbone=Llama2-7B2026.03 | 85.3 | 85.9 | |
| AlphaZeroBudget=Medium, Tokens=2K–4K, Backbone=Llama2-7B2026.03 | 84.3 | 86.2 | |
| ReSCALEBudget=Medium, Tokens=2K–4K, Backbone=Llama2-7B2026.03 | 83.4 | 85.9 | |
| AlphaZeroBudget=Large, Tokens=4K–6K, Backbone=Llama2-7B2026.03 | 82.9 | 84.8 | |
| AlphaZeroBudget=Small, Tokens=0.2K–2K, Backbone=Llama2-7B2026.03 | 74.4 | 86.7 | |
| ReSCALEBudget=Small, Tokens=0.2K–2K, Backbone=Llama2-7B2026.03 | 71.6 | 81.2 | |
| LDDM-GArchitecture=LDDM-G (Ours), Params=85M2025.10 | 63 | — | |
| Best-of-NBudget=N = 32, Tokens=2.5K, Backbone=Llama2-7B2026.03 | 54.1 | — | |
| MGDMArchitecture=MGDM, Params=85M2025.10 | 47 | — | |
| LDDM-GArchitecture=LDDM-G (Ours), Params=6M2025.10 | 28 | — | |
| MGDMArchitecture=MGDM, Params=6M2025.10 | 12 | — |