Mathematical Reasoning on MATH 500 (Token Length Accuracy)
33.8Accuracy (128 tokens)BF16
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| BF16Backbone=LLaDA-8B, Evaluation Protocol=Full-model fine-tuning2026.05 | 33.8 | 38.91 | 40.2 | |
| FlashRL (TIS)Backbone=LLaDA-8B, Evaluation Protocol=Full-model fine-tuning2026.05 | 33.68 | 37.4 | 39 | |
| AISBackbone=LLaDA-8B, Evaluation Protocol=Full-model fine-tuning2026.05 | 32.8 | 38.6 | 40.27 | |
| FP8 RolloutBackbone=LLaDA-8B, Evaluation Protocol=Full-model fine-tuning2026.05 | 30.8 | 37 | 37.8 |