Arithmetic Reasoning on GSM8K (Pass@1, FLOPS)
87.64Pass@1MFS (Ours)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MFS (Ours)Backbone=LLaMA3.1-8B-Instruct2026.01 | 87.64 | — | |
| ϕ-DecodingBackbone=LLaMA3.1-8B-Instruct2026.01 | 86.58 | — | |
| Predictive DecodingBackbone=LLaMA3.1-8B-Instruct2026.01 | 81.43 | — | |
| MCTSBackbone=LLaMA3.1-8B-Instruct2026.01 | 80.44 | — | |
| Tree-of-ThoughtsBackbone=LLaMA3.1-8B-Instruct2026.01 | 75.74 | — | |
| Guided DecodingBackbone=LLaMA3.1-8B-Instruct2026.01 | 75.51 | — | |
| Auto-Regressive (CoT)Backbone=LLaMA3.1-8B-Instruct2026.01 | 70.28 | — | |
| MFS (Ours)Backbone=Mistral-v0.3-7B-Instruct2026.01 | 61.64 | — | |
| ϕ-DecodingBackbone=Mistral-v0.3-7B-Instruct2026.01 | 60.42 | — | |
| MCTSBackbone=Mistral-v0.3-7B-Instruct2026.01 | 60.12 | — | |
| Predictive DecodingBackbone=Mistral-v0.3-7B-Instruct2026.01 | 58 | — | |
| Tree-of-ThoughtsBackbone=Mistral-v0.3-7B-Instruct2026.01 | 53.9 | — | |
| Guided DecodingBackbone=Mistral-v0.3-7B-Instruct2026.01 | 53.9 | — | |
| Auto-Regressive (CoT)Backbone=Mistral-v0.3-7B-Instruct2026.01 | 49.05 | — |