Mathematical Reasoning on GPQA (mean@8)
43.81Mean@8Prompt-GDRO
Evaluation Results
| Method | Links | |
|---|---|---|
| Prompt-GDROBackbone=Qwen3-8B-Base2026.01 | 43.81 | |
| Rollout-GDROBackbone=Qwen3-8B-Base2026.01 | 43.31 | |
| GRPO (Baseline)Backbone=Qwen3-8B-Base2026.01 | 42.25 | |
| Prompt-GDROBackbone=Qwen3-4B-Base2026.01 | 40.88 | |
| Rollout-GDROBackbone=Qwen3-4B-Base2026.01 | 38.51 | |
| GRPO (Baseline)Backbone=Qwen3-4B-Base2026.01 | 35.54 | |
| Prompt-GDROBackbone=Qwen3-1.7B-Base2026.01 | 29.29 | |
| Rollout-GDROBackbone=Qwen3-1.7B-Base2026.01 | 27.72 | |
| GRPO (Baseline)Backbone=Qwen3-1.7B-Base2026.01 | 26.39 |