Mathematical Reasoning on MathVerse (testmini)
71.5AccuracyPRISM + GSPO
Evaluation Results
| Method | Links | |
|---|---|---|
| PRISM + GSPOBase Model=Qwen3-VL-8B2026.04 | 71.5 | |
| PRISM + GRPOBase Model=Qwen3-VL-8B2026.04 | 71.3 | |
| PRISM + DAPOBase Model=Qwen3-VL-8B2026.04 | 70.9 | |
| + DAPOBase Model=Qwen3-VL-8B2026.04 | 69.8 | |
| PRISM + GRPOBase Model=Qwen3-VL-4B2026.04 | 68.6 | |
| PRISM + DAPOBase Model=Qwen3-VL-4B2026.04 | 68.2 | |
| + GRPOBase Model=Qwen3-VL-8B2026.04 | 66.9 | |
| PRISM + GSPOBase Model=Qwen3-VL-4B2026.04 | 66.6 | |
| + GSPOBase Model=Qwen3-VL-8B2026.04 | 65.5 | |
| + DAPOBase Model=Qwen3-VL-4B2026.04 | 65.1 | |
| + GRPOBase Model=Qwen3-VL-4B2026.04 | 64.5 | |
| + GSPOBase Model=Qwen3-VL-4B2026.04 | 64 | |
| InstructBase Model=Qwen3-VL-8B2026.04 | 62.4 | |
| PRISMBase Model=Qwen3-VL-8B2026.04 | 62.2 | |
| + SFTBase Model=Qwen3-VL-8B2026.04 | 60.4 | |
| PRISMBase Model=Qwen3-VL-4B2026.04 | 59.5 | |
| InstructBase Model=Qwen3-VL-4B2026.04 | 59 | |
| + SFTBase Model=Qwen3-VL-4B2026.04 | 58.4 |