Axis Spatial · public evidence index

Geospatial AI
Benchmark Registry

The registry records what each benchmark measures and how results are checked. It also states where the evidence stops. 19 benchmarks are indexed from linked public sources. Unknown fields remain explicit.

19public benchmarks indexed

7own benchmark tasks

11runs recorded

360trials executed

Registry matrix

Registry last reviewed 24 Aug 2026 · horizontal scroll on small screens

In shortEach row is one public benchmark. Exec = agents run real code; Det. verify = answers are checked by recomputation, not by another model. Open? = dataset and code are published so you can verify claims yourself. Follow a benchmark's name for its full record: papers, data, limits, and how to read a score.

Five columns for scanning; every claim keeps its explicit unknown. Each benchmark name opens its full record page.
Benchmark	Tasks	Exec?	Det. verify?	Open?
GeoAnalystBench / GISclaw	50	yes	yes	yes
GeoBenchX	202	yes	no	partial
GEO-Bench-2	19	no	yes	partial
GEO-Bench-2 Leaderboard	—	no	yes	partial
GEOBench-VLM	pending	no	yes	partial
GeoMMBench / GeoMMAgent	1053	no	yes	yes
GABench / GeoAgentBench	53	yes	no	yes
MapQA	3154	no	yes	yes
ThinkGeo	pending	pending verification	yes	no
Earth-Bench / Earth-Agent	pending	pending verification	yes	no
OpenEarth-Bench / OpenEarthAgent	pending	pending verification	yes	no
TerraBench	pending	pending verification	pending verification	no
RSRCC	126000	no	yes	no
UrbanSARFloods	8879	no	yes	no
GeoAI Agency Primitives	pending	yes	yes	no
Axis Synthetic Migration Cases	3	yes	yes	no
GeoNatureAgent Benchmark	93	yes	yes	yes
EO-Gym	9078	yes	yes	yes
MultiGlobeQA	46,060	yes	yes	yes

How to read the registry

Use the JSON endpoint

Comparison unit

A benchmark result is meaningful only with its task set, dataset, tool surface, runtime and verifier. Model names alone are not comparable.

Verification standard

Deterministic checks and inspectable artifacts offer stronger evidence than self-reported scores. Where a benchmark relies on model judgement we say so.

Honest unknowns

Fields marked pending verification could not be confirmed from public sources. They stay visible rather than being guessed.

Arena's own results

The Arena executes its own sealed task pack and publishes every run, including blocked, inconclusive, and rejected outcomes. See the methodology.