Multi-task Language Understanding on MMLU (Acc, AUROC, ECE)
64AccuracyC3RL (Ours)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| C3RL (Ours)Backbone=Llama-3.2-3B-Instruct2026.07 | 64 | 68 | 8.9 | |
| C3RL w/o RefBackbone=Llama-3.2-3B-Instruct2026.07 | 63.5 | 68.8 | 8.7 | |
| RLVRBackbone=Llama-3.2-3B-Instruct2026.07 | 63.4 | 53.7 | 20.8 | |
| RLCRBackbone=Llama-3.2-3B-Instruct2026.07 | 62.3 | 67 | 15.3 | |
| SCBackbone=Llama-3.2-3B-Instruct2026.07 | 59.9 | 73.3 | 15.8 | |
| BaseBackbone=Llama-3.2-3B-Instruct2026.07 | 59.6 | 58.6 | 27.1 | |
| SFT+RefBackbone=Llama-3.2-3B-Instruct2026.07 | 52.8 | 62.8 | 16.2 | |
| SaySelfBackbone=Llama-3.2-3B-Instruct2026.07 | 37.4 | 71.5 | 11 |