Mathematical Reasoning on GSM8K 256
81.58AccuracyThreshold
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| ThresholdModel=LLaDA 1.52026.04 | 81.58 | 72.92 | 65 | |
| FullModel=LLaDA 1.52026.04 | 81.05 | 256 | 18.3 | |
| NI SamplingModel=LLaDA 1.52026.04 | 80.67 | 53.59 | 85.6 | |
| FullModel=LLaDA-8B-Instruct2026.04 | 77.63 | 256 | 18.6 | |
| ThresholdModel=LLaDA-8B-Instruct2026.04 | 77.33 | 74.34 | 62.9 | |
| NI SamplingModel=LLaDA-8B-Instruct2026.04 | 77.18 | 50.97 | 90.2 | |
| FullBackbone=Dream-7B-Base, Sampling Strategy=Full-step2026.04 | 75.05 | 256 | 23 | |
| NI SamplingBackbone=Dream-7B-Base2026.04 | 74.45 | 108.26 | 52.1 | |
| ThresholdBackbone=Dream-7B-Base, Sampling Strategy=Confidence Threshold2026.04 | 72.78 | 161.95 | 36.4 |