Geolocalization on MAPBench easy 1.0 (test)
44.98Acc@500mThinking with Map + Reinforcement Learning + Parallelx4 & Verifier
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Thinking with Map + Reinforcement Learning + Parallelx4 & VerifierBackbone=Qwen3-VL-30B-A3B, Parallel Sampling=x4, Verifier=Yes2026.01 | 44.98 | 55.02 | 80.27 | 82.27 | 85.79 | 91.3 | |
| Thinking with Map + Reinforcement Learning + Parallelx2 & VerifierBackbone=Qwen3-VL-30B-A3B, Parallel Sampling=x2, Verifier=Yes2026.01 | 43.65 | 54.38 | 79.93 | 82.27 | 85.12 | 90.64 | |
| Thinking with Map + Reinforcement LearningBackbone=Qwen3-VL-30B-A3B, Training=RL2026.01 | 41.51 | 50.88 | 76.88 | 79.35 | 83.07 | 89.67 | |
| Thinking with MapBackbone=Qwen3-VL-30B-A3B2026.01 | 33.1 | 40.28 | 53.68 | 56.89 | 59.94 | 64.73 | |
| Gemini-3-ProModel Type=Closed Source, Tools=Google Search/Map2026.01 | 20.86 | 48.28 | 74.31 | 80.69 | 86.9 | 93.79 | |
| Qwen3-VL-235B-A22BModel Type=Open Source2026.01 | 9.35 | 34.06 | 86.14 | 88.48 | 90.82 | 93.66 | |
| GPT-5Model Type=Closed Source2026.01 | 9.02 | 34.39 | 87.48 | 90.32 | 92.99 | 95.49 | |
| GPT-o3Model Type=Closed Source2026.01 | 7.68 | 35.23 | 86.64 | 88.98 | 89.82 | 92.32 | |
| Qwen3-VL-30B-A3BModel Type=Open Source2026.01 | 4.01 | 21.87 | 68.61 | 71.95 | 75.63 | 83.31 | |
| GeoVista-7BModel Type=Open Source, Tools=Google Search2026.01 | 0.33 | 4.17 | 28.21 | 39.39 | 47.74 | 51.08 | |
| GLOBE-7BModel Type=Open Source2026.01 | 0.17 | 6.53 | 42.21 | 58.29 | 73.7 | 82.91 |