Geometric problem solving on Geomverse
70.7AccuracyA²PO
Evaluation results
| Method | Links | ||
|---|---|---|---|
| A²POBase Model=Qwen2.5-VL-7B-Instruct2026.03 | 70.7 | 55.76 | |
| GeometryZeroBase Model=Qwen2.5-VL-7B-Instruct2026.03 | 68.3 | 54.07 | |
| GRPOBase Model=Qwen2.5-VL-7B-Instruct2026.03 | 67.4 | 54.06 | |
| ToRLBase Model=Qwen2.5-VL-7B-Instruct2026.03 | 65.5 | 52.92 | |
| SFTBase Model=Qwen2.5-VL-7B-Instruct2026.03 | 63.6 | 49.02 | |
| GRPOBase Model=Qwen2.5-VL-3B-Instruct2026.03 | 59.1 | 47.01 | |
| ToRLBase Model=Qwen2.5-VL-3B-Instruct2026.03 | 58.4 | 44.71 | |
| A²POBase Model=Qwen2.5-VL-3B-Instruct2026.03 | 58.4 | 48.22 | |
| GeometryZeroBase Model=Qwen2.5-VL-3B-Instruct2026.03 | 57 | 46.35 | |
| SFTBase Model=Qwen2.5-VL-3B-Instruct2026.03 | 56.2 | 39.56 | |
| GeometryZero-VL-7BTraining=GCPO2025.06 | 15.84 | — | |
| Qwen2.5-VL-7B-Instruct + GRPOTraining=GRPO2025.06 | 13.58 | — | |
| Qwen2.5-VL-7B-Instruct + SFTTraining=SFT2025.06 | 12.78 | — | |
| Qwen2.5-VL-7B-InstructBase Model=Qwen2.5-VL-7B-Instruct2025.06 | 11.34 | — | |
| Qwen2.5-VL-7B-Instruct + ToRLTraining=ToRL2025.06 | 9.35 | — |