Reward Modeling on RubricBench
59.7Instruction Following (IF) ScoreEvoRubrics
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| EvoRubricsModel=Qwen3-4B2026.06 | 59.7 | 49.2 | 59.4 | 35 | 49.5 | 51.9 | |
| Deepseek-V4-FlashMethod=Prompt2026.06 | 57.26 | 58.4 | 54.24 | 31.25 | 51.9 | 53.01 | |
| Qwen3-4BMethod=Vanilla2026.06 | 54 | 54 | 50.6 | 28.8 | 47.2 | 48.9 | |
| EvoRubricsModel=Qwen3-8B2026.06 | 50 | 60.8 | 56.09 | 41.25 | 47.87 | 52.4 | |
| Qwen3-32BMethod=Prompt2026.06 | 47.58 | 52 | 55.72 | 40 | 51.9 | 51.53 | |
| Qwen3-8BMethod=SFT2026.06 | 46.77 | 57.6 | 56.46 | 40 | 46.45 | 50.83 | |
| Qwen3-4BMethod=SFT2026.06 | 45.2 | 50 | 59 | 57.5 | 46 | 50.7 | |
| Qwen3-8BMethod=Vanilla2026.06 | 44.35 | 58 | 54.61 | 37.5 | 46.21 | 49.96 |