Logical Reasoning on ZebraLogic (Avg@8)
96.3Accuracy (Avg@8)CERL full
Evaluation Results
| Method | Links | |
|---|---|---|
| CERL fullTraining=SFT+RL2026.05 | 96.3 | |
| HaloTraining=SFT+RL2026.05 | 92 | |
| Fixed-ratio erasure + free answerTraining=None2026.05 | 91.5 | |
| Length-penalty RLTraining=RL2026.05 | 88.7 | |
| Base Qwen3-32BTraining=N/A2026.05 | 88.6 | |
| TokenSkip†Training=SFT2026.05 | 81.2 |