Mathematical Reasoning on MATH 500 (ArithACC, Length)
91.6Arithmetic AccuracyCoLD Base PRM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CoLD Base PRMPolicy Model=Deepseek-v3.12025.07 | 91.6 | 646.2 | |
| CoLD Base PRMPolicy Model=Qwen3-Next-80B-A3B-Instruct2025.07 | 90.4 | 754.5 | |
| Vanilla Base PRMPolicy Model=Deepseek-v3.12025.07 | 89.6 | 953.9 | |
| Vanilla Base PRMPolicy Model=Qwen3-Next-80B-A3B-Instruct2025.07 | 85.2 | 1,050.6 | |
| CoT-SFTBackbone=Qwen3-VL-8B2026.06 | 69.4 | 310.6 | |
| CoT-SFTBackbone=Qwen3-VL-4B2026.06 | 55.8 | 291.5 | |
| DLR (Full)Backbone=Qwen3-VL-8B2026.06 | 54 | 29.5 | |
| DLR (Full)Backbone=Qwen3-VL-4B2026.06 | 44.6 | 28.4 | |
| DLR (SFT)Backbone=Qwen3-VL-8B2026.06 | 39.8 | 27.2 | |
| RoTBackbone=Qwen3-VL-8B2026.06 | 37.8 | 64 | |
| DLR (SFT)Backbone=Qwen3-VL-4B2026.06 | 35.4 | 27.9 | |
| DLR (Full)Backbone=Qwen3-VL-2B2026.06 | 34 | 28.9 | |
| Direct-SFTBackbone=Qwen3-VL-8B2026.06 | 33.5 | 0 | |
| RoTBackbone=Qwen3-VL-4B2026.06 | 33.2 | 64 | |
| Direct-SFTBackbone=Qwen3-VL-4B2026.06 | 29.4 | 0 | |
| DLR (SFT)Backbone=Qwen3-VL-2B2026.06 | 29.4 | 26.7 | |
| CoT-SFTBackbone=Qwen3-VL-2B2026.06 | 29.2 | 324.5 | |
| RoTBackbone=Qwen3-VL-2B2026.06 | 24 | 64 | |
| Direct-SFTBackbone=Qwen3-VL-2B2026.06 | 20.8 | 0 |