Reasoning on AGIEval (Acc, AUROC, ECE)
37.3AccuracyRLVR
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| RLVRBackbone=Llama-3.2-3B-Instruct2026.07 | 37.3 | 0.522 | 0.468 | |
| C3RL (Ours)Backbone=Llama-3.2-3B-Instruct2026.07 | 37.3 | 0.66 | 0.246 | |
| C3RL w/o RefBackbone=Llama-3.2-3B-Instruct2026.07 | 37.2 | 0.652 | 0.218 | |
| BaseBackbone=Llama-3.2-3B-Instruct2026.07 | 34.8 | 0.557 | 0.423 | |
| RLCRBackbone=Llama-3.2-3B-Instruct2026.07 | 34.4 | 0.658 | 0.111 | |
| SFT+RefBackbone=Llama-3.2-3B-Instruct2026.07 | 32.1 | 0.594 | 0.24 | |
| SCBackbone=Llama-3.2-3B-Instruct2026.07 | 28.6 | 0.644 | 0.352 | |
| SaySelfBackbone=Llama-3.2-3B-Instruct2026.07 | 19.4 | 0.611 | 0.072 |