Math Reasoning on MATH500 (Exact-match accuracy Ngen=128, 256, 512)
34.8Exact-match Accuracy (Ngen=128)DISPO
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| DISPObase_model=SFT w/ s1k, algorithm=DISPO2026.02 | 34.8 | 39.2 | 40.2 | 38.1 | |
| DISPObase_model=LLaDA-8B-Instruct, algorithm=DISPO2026.02 | 34 | 39.2 | 40.2 | 37.8 | |
| diffu-GRPObase_model=SFT w/ s1k, algorithm=diffu-GRPO2026.02 | 33.8 | 38.6 | 40.2 | 37.5 | |
| diffu-GRPObase_model=LLaDA-8B-Instruct, algorithm=diffu-GRPO2026.02 | 33.2 | 37.2 | 39.2 | 36.5 | |
| SFT w/ s1kbase_model=SFT w/ s1k2026.02 | 26.2 | 32.6 | 34.8 | 31.2 | |
| LLaDA-8B-Instructbase_model=LLaDA-8B-Instruct2026.02 | 26 | 32.4 | 36.2 | 31.5 |