Model Calibration on MATH, GSM8K, SelfAware, and TruthfulQA combined
0.086ECECARE-GRPO
Evaluation Results
| Method | Links | |
|---|---|---|
| CARE-GRPOBackbone=Qwen2.5-7B2026.01 | 0.086 | |
| RKL-GSPOBackbone=Qwen2.5-7B2026.01 | 0.088 | |
| Base ModelBackbone=Qwen2.5-7B2026.01 | 0.089 | |
| RKL-DAPOBackbone=Qwen2.5-7B2026.01 | 0.09 | |
| RKL-GRPOBackbone=Qwen2.5-7B2026.01 | 0.095 | |
| CARE-DAPOBackbone=Qwen2.5-7B2026.01 | 0.099 | |
| CARE-GSPOBackbone=Qwen2.5-7B2026.01 | 0.101 | |
| GRPO (No Constraint)Backbone=Qwen2.5-7B2026.01 | 0.145 | |
| GSPO (No Constraint)Backbone=Qwen2.5-7B2026.01 | 0.149 | |
| DAPO (No Constraint)Backbone=Qwen2.5-7B2026.01 | 0.151 |