Mathematical Reasoning on GSM8K (Training Time and Inference Time)
295.2Inference TimeBaseline
Evaluation results
| Method | Links | ||
|---|---|---|---|
| BaselineBackbone=Qwen3.5-0.8B, Hardware=NVIDIA A100, Inference Batch Size=2002026.06 | 295.2 | — | |
| Optimized Image (ART)Backbone=Qwen3.5-0.8B, Hardware=NVIDIA A100, Inference Batch Size=2002026.06 | 316.9 | 1,093.6 | |
| Fixed Seed ImageBackbone=Qwen3.5-0.8B, Hardware=NVIDIA A100, Inference Batch Size=2002026.06 | 339 | — | |
| LoRA (DAPO)Backbone=Qwen3.5-0.8B, Hardware=NVIDIA A100, Inference Batch Size=2002026.06 | 354.4 | 2,008.1 |