Reward Prediction on NuminaMath, SIQA, and PIQA (Out-of-Domain Suite)
66.96Average ScoreMemReward
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MemRewardBackbone=Qwen2.5-3B-Instruct2026.03 | 66.96 | 0.89 | |
| R1-OracleBackbone=Qwen2.5-3B-Instruct2026.03 | 66.07 | 0 | |
| R1-pBackbone=Qwen2.5-3B-Instruct2026.03 | 64.44 | -1.63 | |
| MemRewardBackbone=Qwen2.5-1.5B-Instruct2026.03 | 62.81 | 0.81 | |
| R1-OracleBackbone=Qwen2.5-1.5B-Instruct2026.03 | 62 | 0 | |
| R1-pBackbone=Qwen2.5-1.5B-Instruct2026.03 | 58.81 | -3.19 |