Mathematical Reasoning on GSM8K (test) (TPF, TPS, Speedup, Solve Rate)
1TPF (Time Per First Token)AR
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| ARFamily=AR-based, Backbone=Qwen2.5-Math-7B-Instruct2025.12 | 1 | 41.8 | 1 | 92.4 | |
| LLaDA-InstructFamily=Diffusion-based2025.12 | 1 | 7.2 | 0.17 | 77.4 | |
| Dream-BaseFamily=Diffusion-based2025.12 | 1 | 9.5 | 0.23 | 75 | |
| JacobiFamily=AR-based, Backbone=Qwen2.5-Math-7B-Instruct2025.12 | 1.05 | 42.2 | 1.02 | 92.4 | |
| Fast-dLLM (DC)Family=Diffusion-based2025.12 | 2.1 | 49.8 | 1.19 | 75 | |
| CLLM*Family=AR-based, Backbone=Qwen2.5-Math-7B-Instruct2025.12 | 2.25 | 86.8 | 2.08 | 92.2 | |
| D2FFamily=Diffusion-based2025.12 | 3.1 | 91.2 | 2.18 | 77.6 | |
| Jacobi Forcing ModelFamily=AR-based, Backbone=Qwen2.5-Math-7B-Instruct2025.12 | 3.72 | 146.1 | 3.5 | 91.4 | |
| Jacobi Forcing Model (MR)Family=AR-based, Backbone=Qwen2.5-Math-7B-Instruct, Optimization=Multi-block Decoding with Rejection recycling2025.12 | 4.04 | 154.9 | 3.71 | 91.4 |