Question Answering on GPQA Diamond (Accuracy and Tokens)
51.5AccuracySFT + RL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SFT + RLBackbone=Qwen3-8B2026.04 | 51.5 | 1,382 | |
| Abstract-CoT (Warm-up + RL)Backbone=Qwen3-8B2026.04 | 50.5 | 174 | |
| SFT (CoT)Backbone=Qwen3-8B2026.04 | 45.5 | 1,106 | |
| BaselineBackbone=Qwen3-8B2026.04 | 44.9 | 767 | |
| Abstract-CoT (Warm-up)Backbone=Qwen3-8B2026.04 | 44.4 | 168 | |
| Abstract-CoT (RL-only)Backbone=Qwen3-8B2026.04 | 41.9 | 142 | |
| SFT (no CoT)Backbone=Qwen3-8B2026.04 | 39.4 | 499 |