Expert-level Question Answering on GPQA-D (Accuracy @8)
68.9Accuracy @8CERL full
Evaluation results
| Method | Links | |
|---|---|---|
| CERL fullTraining=SFT+RL2026.05 | 68.9 | |
| HaloTraining=SFT+RL2026.05 | 66.3 | |
| Fixed-ratio erasure + free answerTraining=None2026.05 | 66.1 | |
| Base Qwen3-14BTraining=N/A2026.05 | 64.1 | |
| Length-penalty RLTraining=RL2026.05 | 63.3 | |
| TokenSkip†Training=SFT2026.05 | 58.1 |