Datasets:
Tasks:
Question Answering
Modalities:
Geospatial
Languages:
English
Size:
10K<n<100K
ArXiv:
License:
| license: mit | |
| task_categories: | |
| - question-answering | |
| language: | |
| - en | |
| tags: | |
| - geospatial | |
| - gps | |
| - benchmark | |
| - geography | |
| - spatial-reasoning | |
| - coordinates | |
| size_categories: | |
| - 10K<n<100K | |
| pretty_name: "GPSBench: GPS Reasoning Benchmark for LLMs" | |
| dataset_info: | |
| - config_name: pure_gps | |
| description: "Pure GPS track: coordinate manipulation tasks" | |
| - config_name: applied | |
| description: "Applied track: geographic reasoning tasks" | |
| # GPSBench: Do Large Language Models Understand GPS Coordinates? | |
| GPSBench is a benchmark dataset of **57,800 samples** across **17 tasks** for evaluating geospatial reasoning in Large Language Models (LLMs). | |
| - **Paper**: [arXiv:2602.16105](https://arxiv.org/abs/2602.16105) | |
| - **Code**: [github.com/joey234/gpsbench](https://github.com/joey234/gpsbench) | |
| - **Leaderboard**: [gpsbench.github.io](https://gpsbench.github.io) | |
| ## Benchmark Structure | |
| GPSBench is organized into two complementary evaluation tracks: | |
| ### Pure GPS Track (9 tasks) | |
| Coordinate manipulation without geographic knowledge: | |
| - **Representation**: Format Conversion, Coordinate System Transformation | |
| - **Measurement**: Distance Calculation, Bearing Computation, Area & Perimeter | |
| - **Spatial Operations**: Coordinate Interpolation, Bounding Box, Route Geometry, Relative Position | |
| ### Applied Track (9 tasks) | |
| Real-world geographic reasoning requiring world knowledge: | |
| - **Knowledge Retrieval**: Place Association, Name Disambiguation, Terrain Classification | |
| - **Spatial Reasoning**: Relative Position, Proximity & Nearest Neighbor, Boundary Analysis | |
| - **Pattern Analysis**: Route Analysis, Spatial Patterns, Missing Data Inference | |
| ## Dataset Splits | |
| | Split | Ratio | Samples | | |
| |-------|-------|---------| | |
| | Train | 60% | ~34,680 | | |
| | Dev | 10% | ~5,780 | | |
| | Test | 30% | ~17,340 | | |
| Each task contains approximately 3,400 samples (1,020 test). | |
| ## Data Format | |
| Each sample is a JSON object containing: | |
| - `task`: Task identifier | |
| - `question`: Human-readable question/prompt | |
| - `ground_truth`: Expected answer with evaluation metrics | |
| - `coordinate(s)`: GPS coordinate data | |
| - `metadata`: Track, task description, and other context | |
| ## Citation | |
| ```bibtex | |
| @article{gpsbench2025, | |
| title = {GPSBench: Do Large Language Models Understand GPS Coordinates?}, | |
| author = {Truong, Thinh Hung and Lau, Jey Han and Qi, Jianzhong}, | |
| journal = {arXiv preprint arXiv:2602.16105}, | |
| year = {2025} | |
| } | |
| ``` | |