cs.CVApr 2, 2026

GeoAI Agency Primitives

Authors: Akram Zaytar, Rohan Sawahn, Caleb Robinson, Gilles Q. Hacheme, Girmaw A. Tadesse, Inbal Becker-Reshef, Rahul Dodhia, Juan Lavista Ferres

Organizations: Microsoft AI for Good Lab · NASA Harvest

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

We present ongoing research on agency primitives for GeoAI assistants -- core capabilities that connect Foundation models to the artifact-centric, human-in-the-loop workflows where GIS practitioners actually work. Despite advances in satellite image captioning, visual question answering, and promptable segmentation, these capabilities have not translated into productivity gains for practitioners who spend most of their time producing vector layers, raster maps, and cartographic products. The gap is not model capability alone but the absence of an agency layer that supports iterative collaboration. We propose a vocabulary of 99 primitives for such a layer -- including navigation, perception, geo-referenced memory, and dual modeling -- along with a benchmark that measures human productivity. Our goal is a vocabulary that makes agentic assistance in GIS implementable, testable, and comparable.

Figures & tables

Figure 1: General agent workflow. User queries are parsed, decomposed into tasks, and executed recursively. Completed tasks yield suggestions for user review before committing.
Figure 2: Core sensing primitives. (Top-left) Navigation constructs context bundles specifying sub-ROIs, zoom, and sampling strategy. (Top-right) Perception routes patches to task-appropriate models, returning labels and notes. (Bottom-left) GeoMemory stores spatial notes for retrieval and curation. (Bottom-right) Embeddings map inputs to vectors for similarity search and modeling.
Figure 3: Execution and enrichment primitives. (Top-left) Compute Graphs translate queries into directed operation graphs. (Top-right) Budgets enforce constraints enabling partial results and early stopping. (Bottom-left) Attribution enriches geometries with external data. (Bottom-right) Dual Modeling iterates between expensive VLM judgments and cheap scalable inference.
MetricDefinition
Time-to-thresholdTτ=min⁡{t:Q(t)≥τ}T_{\tau}=\min\{t:Q(t)\geq\tau\} , time to reach quality τ\tau
Progress AUC1T∫0TQ(t) dt\frac{1}{T}\int_{0}^{T}Q(t)\,dt , normalized area under quality curve
Rework rateR=noverwrite/neditsR=n_{\text{overwrite}}/n_{\text{edits}} , fraction of edits that revise prior work
Suggestion bias∥Perraccept−Perrgt∥\|P_{\text{err}}^{\text{accept}}-P_{\text{err}}^{\text{gt}}\| , error distribution shift from suggestions
Table 1: The benchmark tracks four primary metrics. In addition, accept/reject rates, compute cost, and GIS validity (geometry, CRS, schema) are logged for further evaluation.
Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

SystemApproachSuggest-ReviewGeoMemoryBudgetsPropag.
GeoGPT ( Zhang et al., 2023 )Tool-call————
AutoGIS ( Li and Ning, 2023 )Code-gen————
MapGPT ( Zhang et al., 2024 )Tool-callpartial———
Tree-GPT ( Du et al., 2023 )Hybrid————
GIS Copilot ( Akinboyewa et al., 2025 )Tool+codepartial———
GeoAgent ( Chen et al., 2024 )Code+MCTS——partial—
Appendix
Table 2: Capability comparison of GeoAI agent frameworks. Existing systems span three paradigms: Tool-calling uses LangChain-style agents with predefined function pools; Code-generation has LLMs produce Python scripts directly; Hybrid systems combine multiple approaches. Columns indicate: Suggest-Review (iterative human approval before commit), Georef. Memory (spatially-indexed state across sessions), Budgets (constraints on compute/time/cost), Propagation (“find more like this” for efficient labeling).
More in this topic

Agentic Workflows

Agentic workflows coordinate AI agents, tools, and humans across multi-step tasks, emphasizing reliable execution, efficiency, and oversight.

See all 218 papers

Explore similar work

CardsList

  1. Google Earth AI Geospatial Reasoning Agent: Capabilities and a Systematic Evaluation

    Oct 11, 2026Noam Slonim, Aviv Slobodkin, Gil Fidel +34Geospatial QAGeospatial Reasoning

    Google Research · Google Geo · Spatial Bytes

  2. GISAgentBench: A Practitioner-Sourced Benchmark for Evaluating LLM Agents on GIS Tasks

    Aug 3, 2026Abhinav Pothuri, Zhe Jiang, Zelin Xu +1LLM Agent EvaluationAI Agent Benchmarks

    Department of Computer & Information Science & Engineering, University of Florida · Department of Geography, University of Florida

  3. GISclaw: A Comprehensive Open-Source LLM Agent System for Realistic Multi-Step Geospatial Analysis

    Mar 27, 2026Jinzhen Han, JinByeong Lee, Yuri Shim +2Geospatial Reasoning BenchmarksGeospatial Reasoning

    Department of Civil, Architectural & Environment Engineering, Sungkyunkwan University, Suwon, South Korea · Ministry of the Interior and Safety, South Korea · School of Geography, University of Leeds, United Kingdom +1