Mathematical Reasoning on GSM8K (Random, Reward, ∆ (pp))
90.4Random BaselineSparse Reward
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Sparse RewardBackbone=Qwen3-4B2025.10 | 90.4 | 93 | 2.5 | |
| Dense RewardBackbone=Qwen3-4B2025.10 | 89.6 | 91.4 | 1.8 | |
| Interval RewardBackbone=Qwen3-4B2025.10 | 87.8 | 91.2 | 3.5 | |
| Sparse RewardBackbone=Qwen2.5-7B2025.10 | 85.8 | 88.8 | 3 | |
| Sparse RewardBackbone=Llama3.1-8B2025.10 | 80.6 | 82.1 | 1.5 | |
| Interval RewardBackbone=Qwen2.5-7B2025.10 | 78.8 | 82.5 | 3.7 | |
| Interval RewardBackbone=Llama3.1-8B2025.10 | 67.6 | 78.2 | 10.7 | |
| Dense RewardBackbone=Llama3.1-8B2025.10 | 64.6 | 71.1 | 6.5 | |
| Dense RewardBackbone=Qwen2.5-7B2025.10 | 38.4 | 55.7 | 17.4 |