Spatial reasoning
Measure whether a model can interpret geographic context, constraints, relationships and scale across realistic regional tasks.
Solutions · Model Evaluation
Measure how well AI
understands territory.
LGM evaluates existing and future models on real geospatial and city-planning tasks, then turns failures into a practical roadmap for better tools, data, evaluation and regional adaptation.
Evaluation scope
Beyond text quality.
Test the spatial system.
A geospatial answer is useful only when the data, operation, geometry and evidence are correct. The benchmark evaluates the complete path.
Planning & risk scenarios
Evaluate urban planning, territorial development and infrastructure, environmental and climate-risk decisions.
Tool-call correctness
Test whether the system chooses the right geospatial operation, supplies valid parameters and handles tool failure.
GIS-ready outputs
Score spatial calculations, geometry, attributes, evidence quality, provenance and reproducibility—not only fluent text.
R&D workflow
A reproducible baseline.
A focused improvement plan.
A bounded delivery path keeps the work measurable, reviewable and ready to move from validation into production.
- 01
Define the baseline
Choose target users, regions, decisions, tools and objective success criteria.
- 02
Run the benchmark
Execute a versioned task set with traceable sources, tool calls, calculations and outputs.
- 03
Diagnose failures
Separate reasoning, retrieval, data, tool selection, geometry and presentation errors.
- 04
Plan improvement
Prioritize better tools, data, prompts, evaluation, fine-tuning and regional adaptation.
Benchmark outputs
Evidence for the next
model decision.
The engagement leaves the partner with reusable evaluation assets, not only a one-time score.
- Analytical capability report
- Versioned and reproducible benchmark
- Current model and system baseline
- Failure taxonomy with trace examples
- Recommendations for data, tools and fine-tuning
- Roadmap for a specialized geospatial version
Commercial model
Paid R&D with reusable outcomes.
Scope and pricing depend on regions, domains, models, tool surface and publication requirements. Rights to benchmark tasks, datasets and public results are agreed explicitly.
- 01Benchmark design and task authoring
- 02Evaluation runs and analysis
- 03Model and agent comparison
- 04Improvement and adaptation roadmap
- 05Optional recurring regression testing
Representative benchmarks
Evaluate decisions
that depend on place.
The strongest task set comes from real partner or client workflows with inspectable data and a clear definition of success.
City planning copilots
Test zoning, accessibility, service gaps, site comparison and evidence-backed recommendations.
Regional risk agents
Measure infrastructure, terrain, climate and environmental risk analysis across real territories.
Model and agent platforms
Compare models and agent configurations before publishing a specialized regional or geospatial version.
Geospatial & City Planning Benchmark