Math Reasoning on GSM8K (Ngen Exact Match Accuracy)
75Exact Match Accuracy (Ngen=128)DISPO
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| DISPObase_model=SFT w/ s1k, algorithm=DISPO2026.02 | 75 | 82.3 | 83.2 | 80.2 | |
| DISPObase_model=LLaDA-8B-Instruct, algorithm=DISPO2026.02 | 74.2 | 80.6 | 81.8 | 78.9 | |
| diffu-GRPObase_model=SFT w/ s1k, algorithm=diffu-GRPO2026.02 | 73.2 | 81.1 | 82.1 | 78.8 | |
| diffu-GRPObase_model=LLaDA-8B-Instruct, algorithm=diffu-GRPO2026.02 | 72.6 | 79.8 | 81.9 | 78.1 | |
| LLaDA-8B-Instructbase_model=LLaDA-8B-Instruct2026.02 | 68.7 | 76.7 | 78.2 | 74.5 | |
| SFT w/ s1kbase_model=SFT w/ s1k2026.02 | 66.5 | 78.8 | 81.1 | 75.5 |