GeoAgentBench (Plan-and-React) - Map Verification: leaderboard
GeoAgentBench (2026) is a dynamic execution benchmark for tool-augmented GIS agents: spatial analysis tasks run in a sandbox of 117 atomic GIS tools and are scored on the tool trajectory, the parameters of the executed calls and the produced map. This board is the 0 to 100 score a GPT-4o judge gives the agent's final map against the gold map for spatial accuracy and cartographic style, with the agent using a global plan with step-wise reactive execution; higher is better.
Metric: VLM verification score (0 to 100, mean of three GPT-4o judgements) comparing the agent's final map with the gold-trajectory map for data-spatial accuracy and cartographic style, on 53 spatial analysis tasks over 117 atomic GIS tools in an executable sandbox (30 steps per task), agents under the authors' Plan-and-React framework (global plan with step-wise reactive execution); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 79 | #173 |
| 2 | DeepSeek V3 | 68.5 | #729 |
| 3 | GPT-4o | 59.1 | #751 |
| 4 | Gemini 2.5 Flash | 52.9 | #463 |
| 5 | GPT-4o Mini | 43.4 | #1667 |
| 6 | Qwen 2.5 7B Instruct | 20.1 | #2723 |
| 7 | Llama 3.1 8B Instruct | 10.1 | #4600 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=geoagentbench-plan-and-react-map-verification · How It Works · Data refreshed daily, snapshot 2026-11-21.