Mathematical Reasoning on MATH 500 (Accuracy and Latency)
76.2Accuracy (pass@1)Power-SMC
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Power-SMCModel=Qwen2.5-Math-7B2026.02 | 76.2 | 1.44 | |
| Scalable Power SamplingModel=Qwen2.5-Math-7B2026.02 | 75.8 | — | |
| MH power samplingModel=Qwen2.5-Math-7B2026.02 | 74.8 | 18.2 | |
| Power-SMCModel=Qwen2.5-7B2026.02 | 71.4 | 1.64 | |
| Scalable Power SamplingModel=Qwen2.5-7B2026.02 | 70.8 | — | |
| MH power samplingModel=Qwen2.5-7B2026.02 | 70.6 | 28.3 | |
| Low-temperature decodingModel=Qwen2.5-Math-7B, temperature=τ=1/α2026.02 | 69 | 1 | |
| Low-temperature decodingModel=Qwen2.5-7B, temperature=τ=1/α2026.02 | 62.8 | 1.24 | |
| Power-SMCModel=Phi-3.5-mini-instruct2026.02 | 51.6 | 3.25 | |
| MH power samplingModel=Phi-3.5-mini-instruct2026.02 | 50.8 | 16.12 | |
| Baseline decodingModel=Qwen2.5-7B2026.02 | 49.8 | 1 | |
| Baseline decodingModel=Qwen2.5-Math-7B2026.02 | 49.6 | 1 | |
| Low-temperature decodingModel=Phi-3.5-mini-instruct, temperature=τ=1/α2026.02 | 47.8 | 0.91 | |
| Baseline decodingModel=Phi-3.5-mini-instruct2026.02 | 40 | 1 |