Superpower Daily

Superpower Daily

Your reading space

Sign in
Swipe

Research investigation R1014 / benchmark analysis

Google’s geospatial tests show larger score gains from specialized capabilities than added planning and guidance

The published comparisons support descriptive gains from a specialized-capability package, but do not isolate skill packaging, planning alone or a planning-by-skills interaction.

Current public editionv1Oct 15, 2026
Read the original analysis
Verified observations
16

2 measured fields

Supported claims
9

9 material findings

Cited sources
5

5 primary or authoritative

Research score
77

Automated topic and evidence score

Interactive figureGoogle’s geospatial tests show larger score...
CSV JSON
Data status8 verified observations

Bar height represents full_minus_skills. Colors separate the recorded series.

Verified observationHover or focus any mark for exact valuesLast updated Oct 15, 2026

Version ledger

Frozen public editions

Each edition preserves the records, method, sources, and downloads available at publication time.

  1. v1 / latestOct 15, 202616 records / 5 sources

    Initial public snapshot with 16 records and 5 cited sources.

Coverage note

The bounded sample is Google's four configurations across four benchmarks for two models, with aggregate results across three agent runs. Reported per-run test counts are 76 GeoExpertQueries, 220 GeoKTools, 44 adapted CBGB and 140 UnivEARTH queries; the repository instead lists 210 GeoKTools test queries. This component comparison differs from the supplied recent launch headline about benchmark publication and limited early access.

Dataset ID
spd:skills-or-planning-which-adds-more-in-google-s-geospatial-agent-tests-97549bcc
Stable URL
/research/skills-or-planning-which-adds-more-in-google-s-geospatial-agent-tests-97549bcc
Version
v1
Coverage
Live collection
Records
16
Fields
7
Updated

Read the data

The records behind the figure

CSV JSON
Google’s geospatial tests show larger score gains from specialized capabilities than added planning and guidance data records
EntityMetricValueUnitObservedSourceTransform
Gemini 3.1 Pro / CBGBfull_minus_skills0score_points—https://arxiv.org/html/2610.15750v184.8 − 84.8 = 0.0
Gemini 3.1 Pro / GeoExpertQueriesfull_minus_skills-3.2score_points—https://arxiv.org/html/2610.15750v147.6 − 50.8 = −3.2
Gemini 3.1 Pro / GeoKToolsfull_minus_skills-2.9score_points—https://arxiv.org/html/2610.15750v181.6 − 84.5 = −2.9
Gemini 3.1 Pro / UnivEARTHfull_minus_skills-8.4score_points—https://arxiv.org/html/2610.15750v189.5 − 97.9 = −8.4
Gemini 3.8 Flash / CBGBfull_minus_skills-1.5score_points—https://storage.googleapis.com/gweb-research2023-media/images/EarthAITechReport1_table2.width-1250.png86.4 − 87.9 = −1.5
Gemini 3.8 Flash / GeoExpertQueriesfull_minus_skills10.5score_points—https://storage.googleapis.com/gweb-research2023-media/images/EarthAITechReport1_table2.width-1250.png62.7 − 52.2 = 10.5
Gemini 3.8 Flash / GeoKToolsfull_minus_skills-0.1score_points—https://storage.googleapis.com/gweb-research2023-media/images/EarthAITechReport1_table2.width-1250.png85.9 − 86.0 = −0.1
Gemini 3.8 Flash / UnivEARTHfull_minus_skills-0.7score_points—https://storage.googleapis.com/gweb-research2023-media/images/EarthAITechReport1_table2.width-1250.png98.1 − 98.8 = −0.7
Gemini 3.1 Pro / CBGBskills_minus_basic71.2score_points—https://arxiv.org/html/2610.15750v184.8 − 13.6 = 71.2
Gemini 3.1 Pro / GeoExpertQueriesskills_minus_basic48.5score_points—https://arxiv.org/html/2610.15750v150.8 − 2.3 = 48.5
Gemini 3.1 Pro / GeoKToolsskills_minus_basic46.7score_points—https://arxiv.org/html/2610.15750v184.5 − 37.8 = 46.7
Gemini 3.1 Pro / UnivEARTHskills_minus_basic1.2score_points—https://arxiv.org/html/2610.15750v197.9 − 96.7 = 1.2
Gemini 3.8 Flash / CBGBskills_minus_basic68.2score_points—https://storage.googleapis.com/gweb-research2023-media/images/EarthAITechReport1_table2.width-1250.png87.9 − 19.7 = 68.2
Gemini 3.8 Flash / GeoExpertQueriesskills_minus_basic48.9score_points—https://storage.googleapis.com/gweb-research2023-media/images/EarthAITechReport1_table2.width-1250.png52.2 − 3.3 = 48.9
Gemini 3.8 Flash / GeoKToolsskills_minus_basic46score_points—https://storage.googleapis.com/gweb-research2023-media/images/EarthAITechReport1_table2.width-1250.png86.0 − 40.0 = 46.0
Gemini 3.8 Flash / UnivEARTHskills_minus_basic0.5score_points—https://storage.googleapis.com/gweb-research2023-media/images/EarthAITechReport1_table2.width-1250.png98.8 − 98.3 = 0.5

Measurement technique

How to read this report

  1. 01Use only the supplied saved evidence, retaining the October 14, 2026, 20:31:06.926 UTC cutoff. This synthesis is not a new collection or experiment; no independent rerun was performed.
  2. 02Compare standalone, basic-tooling, skills and full configurations within Gemini 3.8 Flash and Gemini 3.1 Pro, separately for each benchmark. The report states that queries, decoding defaults and execution limits match within each model.
  3. 03Calculate skills-minus-basic and full-minus-skills differences from published rounded means. Interpret the former as a specialized-capability package and the latter as planning plus execution guidance.
  4. 04Plan an evidence_matrix with one row per model and benchmark, recording both contrasts, source, scoring definition, reported denominator, uncertainty and attribution boundary. Keep different benchmark outcomes separate; calculate no cross-benchmark average.
  5. 05Classify package-level descriptive comparisons as supported, planning-alone attribution and factorial interaction as unsupported by the configuration design, and statistical significance of component differences as inconclusive.

Sources

Evidence

5 publishers supporting 16 records. Expand a publisher to inspect its cited pages.

arxiv.orgarxiv.org1 source / 8 records
arxiv.org8 supporting records
storage.googleapis.comstorage.googleapis.com1 source / 8 records
storage.googleapis.com8 supporting records
GitHubgithub.com1 source / 0 records
GitHub - google-research/geospatial-reasoning-benchmarks0 supporting records
raw.githubusercontent.comraw.githubusercontent.com1 source / 0 records
raw.githubusercontent.com0 supporting records
research.googleresearch.google1 source / 0 records
Putting Earth AI to work: Geospatial agents for our planet's biggest challenges0 supporting records
Next report / 01AI Model Economics Index All research reports
YOUR READING SPACE

Notifications