Reward Modeling on Overall 5-Benchmark Suite
73.5Average ScoreConsistRM (Qwen3-8B)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ConsistRM (Qwen3-8B)GT=False2026.04 | 73.5 | 2.6 | |
| Qwen3-8B + RFTGT=True2026.04 | 72.5 | 1.6 | |
| Qwen3-8B + CAARGT=False2026.04 | 72.2 | 1.3 | |
| Qwen3-8B + TTRLGT=False2026.04 | 71.7 | 0.8 | |
| ConsistRM (Qwen3-4B)GT=False2026.04 | 71 | 2.4 | |
| Qwen3-8B2026.04 | 70.9 | — | |
| Qwen3-8B + SFTGT=True2026.04 | 70.7 | -0.2 | |
| Qwen3-4B + CAARGT=False2026.04 | 70.3 | 1.7 | |
| Qwen3-4B + RFTGT=True2026.04 | 69.8 | 1.2 | |
| Qwen3-4B + TTRLGT=False2026.04 | 69.5 | 0.9 | |
| Qwen3-4B2026.04 | 68.6 | — | |
| Qwen3-4B + SFTGT=True2026.04 | 67.8 | -0.8 |