Scientific Reasoning on GPQA Diamond (Accuracy Avg@8)
73.4Accuracy (Avg@8)CERL full
Evaluation results
| Method | Links | |
|---|---|---|
| CERL fullTraining=SFT+RL2026.05 | 73.4 | |
| HaloTraining=SFT+RL2026.05 | 70.2 | |
| Fixed-ratio erasure + free answerTraining=None2026.05 | 69 | |
| Base Qwen3-32BTraining=N/A2026.05 | 68.1 | |
| Length-penalty RLTraining=RL2026.05 | 66.2 | |
| TokenSkip†Training=SFT2026.05 | 61.3 |