Mathematical Reasoning on AIME 24 (Accuracy, Training Tokens, TPSPD)
75.8AIME24 AccuracyAsync (ours)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Async (ours)Micro-BS=1, NPUs=16, Global Batch Size=32, Steps=100, Execution=Async2025.11 | 75.8 | 917.609 | 192.259 | |
| VERL (async)Micro-BS=16, NPUs=16, Global Batch Size=32, Steps=100, Execution=async, Backend=FSDP2025.11 | 74.6 | 812.084 | 155.521 | |
| Sync (ours)Micro-BS=1, NPUs=16, Global Batch Size=32, Steps=100, Execution=Sync2025.11 | 74.6 | 827.306 | 99.966 | |
| MindSpeed-RLMicro-BS=1, NPUs=16, Global Batch Size=32, Steps=1002025.11 | 73.3 | 706.355 | 61.641 | |
| Base model (8B)Model size=8B2025.11 | 72.5 | — | — |