Mathematical Reasoning on GSM8K (Accuracy, Latency, Throughput, Length/Steps)
843.05Throughput (TPS)T3D
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| T3DDecoding Strategy=Dynamic, Block Size=4, Steps Per Block=4, Confidence Threshold=0.9, Temperature=0.1, Model=SDAR-4B-Chat2026.02 | 843.05 | 0.37 | 83.03 | 312.48 | 72.4 | |
| dParallelDecoding Strategy=Dynamic, Block Size=4, Steps Per Block=4, Confidence Threshold=0.9, Temperature=0.1, Model=SDAR-4B-Chat2026.02 | 805.02 | 0.39 | 83.23 | 310.58 | 67.02 | |
| Naive TDDecoding Strategy=Dynamic, Block Size=4, Steps Per Block=4, Confidence Threshold=0.9, Temperature=0.1, Model=SDAR-4B-Chat2026.02 | 696.99 | 0.47 | 89.78 | 330.82 | 62.4 | |
| ReDiDecoding Strategy=Dynamic, Block Size=4, Steps Per Block=4, Confidence Threshold=0.9, Temperature=0.1, Model=SDAR-4B-Chat2026.02 | 636.58 | 0.49 | 84.63 | 311.99 | 54.89 | |
| OriginalDecoding Strategy=Dynamic, Block Size=4, Steps Per Block=4, Confidence Threshold=0.9, Temperature=0.1, Model=SDAR-4B-Chat2026.02 | 580.6 | 0.43 | 71.12 | 249.52 | 61.56 |