Mathematical Reasoning on AIME '25 (Accuracy, Tokens)
25.6AccuracySFT + RL
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SFT + RLBackbone=Qwen3-8B2026.04 | 25.6 | 9,343 | |
| Abstract-CoT (Warm-up + RL)Backbone=Qwen3-8B2026.04 | 24.4 | 3,438 | |
| BaselineBackbone=Qwen3-8B2026.04 | 23.3 | 3,981 | |
| SFT (CoT)Backbone=Qwen3-8B2026.04 | 23.3 | 4,627 | |
| Abstract-CoT (Warm-up)Backbone=Qwen3-8B2026.04 | 20 | 2,195 | |
| SFT (no CoT)Backbone=Qwen3-8B2026.04 | 18.9 | 2,745 | |
| Abstract-CoT (RL-only)Backbone=Qwen3-8B2026.04 | 18.9 | 2,849 |