Geospatial Agentic Tool-use and Grounded Reasoning on OpenEarthAgent
80.56Tool Any OrMAS+SFT+GRPO
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| MAS+SFT+GRPOType=Multi2026.06 | 80.56 | 75.56 | 93.33 | 85.16 | 94.39 | 99.8 | 96.76 | 100 | 95.45 | 86.8 | |
| MAS+SFTType=Multi2026.06 | 73.89 | 65.56 | 85 | 80.24 | 85.46 | 99.57 | 96.92 | 100 | 94.7 | 81.17 | |
| OpenEarthAgentType=Multi2026.06 | 67.75 | 67.24 | 72.71 | 45.26 | 75.86 | 99.51 | 97.18 | 96.08 | 62.1 | 83.64 | |
| GPT-4oType=Single2026.06 | 50.81 | 50.38 | 55.52 | 39.22 | 77.93 | 99.18 | 93.88 | 85.48 | 45.8 | 86.76 | |
| GPT-5Type=Single2026.06 | 46.96 | 46.79 | 47.81 | 43.88 | 46.21 | 97.54 | 82.69 | 73.28 | 37.28 | 87.02 | |
| o4-miniType=Single2026.06 | 40.12 | 39.95 | 41.49 | 35.18 | 55.17 | 83.68 | 68.33 | 64.17 | 37.95 | 89.48 | |
| Llama-3.1-8B-Inst.Type=Single2026.06 | 39.01 | 37.72 | 44.91 | 12.7 | 55.17 | 47.07 | 39.3 | 34.11 | 17.49 | 79.08 | |
| Qwen2.5-7B-Inst.Type=Single2026.06 | 31.57 | 30.02 | 36.61 | 15.85 | 41.38 | 94.08 | 85.51 | 78.46 | 38 | 80.08 | |
| Qwen3-4B-Inst.Type=Single2026.06 | 16 | 14.71 | 21.47 | 13.72 | 15.86 | 97.34 | 86.94 | 84.12 | 33.55 | 83.28 | |
| Qwen2.5-3B-Inst.Type=Single2026.06 | 9.24 | 8.72 | 12.4 | 1.83 | 24.14 | 85.07 | 72.13 | 64.87 | 24.12 | 68.75 | |
| Mistral-7B-Inst.Type=Single2026.06 | 2.65 | 2.57 | 2.91 | 0.39 | 5.52 | 64.14 | 35.12 | 26.11 | 12.45 | 24.04 | |
| InternLM3-8B-Inst.Type=Single2026.06 | 2.22 | 1.88 | 3.16 | 0.24 | 2.76 | 44.54 | 38.46 | 27.82 | 13.25 | 29.84 |