pith. sign in

arxiv: 2606.12821 · v1 · pith:QP6NMCGEnew · submitted 2026-06-11 · 💻 cs.AI · cs.ET

GeoNatureAgent Benchmark: Benchmarking LLM Agents for Environmental Geospatial Analysis Across Frontier and Open-Weight Foundation Models

Gabriel Diaz-Ireland , Diego Prieto-Herr\'aez , Mario Garc\'ia Peces , Javier Vel\'azquez , Devika Jain This is my paper

Pith reviewed 2026-06-27 07:16 UTC · model grok-4.3

classification 💻 cs.AI cs.ET
keywords GeoNatureAgent BenchmarkLLM agentsgeospatial analysisenvironmental indicatorstool callingbenchmarkingSpain Portugalcomparison tasks
0
0 comments X

The pith

Claude Sonnet 4 reaches 60.8% success on the first benchmark of LLM agents using real geospatial APIs, while all models score 0% on close-value comparisons.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper creates the GeoNatureAgent Benchmark to measure how well LLM agents can automate environmental geospatial work by calling structured tools on a production-style API that serves real indicators for Spain and Portugal. It runs 93 tasks across categories such as municipality analysis, spatial reasoning, multi-turn conversations, error recovery, and cross-indicator synthesis, then scores seven models under controlled conditions. The results place Claude Sonnet 4 first at 60.8 percent, DeepSeek V3.2 second at 56.3 percent with far lower cost, and every other model below 51 percent. All models fail completely on tasks that require distinguishing close numerical values. The benchmark is shown to be stricter than earlier GIS tests by 25 to 35 points, and the authors demonstrate it can incorporate new data sources without redesign.

Core claim

The GeoNatureAgent Benchmark shows that current LLM agents operating through structured tool calls against a real environmental API achieve at most 60.8 percent success across 93 tasks, with open-weight models occupying most of the cost-accuracy frontier and every model failing entirely on close-value comparison problems.

What carries the argument

A 93-task benchmark that drives agents through sixteen production tools against a self-hostable API serving CO2, erosion, and land-cover indicators across Spain and Portugal.

If this is right

  • Open-weight models can deliver 93 percent of the top model's accuracy at roughly one-eleventh the cost per task.
  • Comparison and ranking tasks that involve near-identical values remain unsolved by all current agents.
  • Structured tool calling on a real API produces lower and more differentiated scores than general-purpose GIS benchmarks.
  • The benchmark framework can be extended by adding new indicator layers without changing the task or evaluation design.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Specialized numerical-comparison modules may be required before agents can handle the ranking and difference tasks that currently block all models.
  • Environmental analysis pipelines could combine LLM agents with separate deterministic comparison tools to bypass the observed zero-success regime.
  • Lower-cost open models become practical for routine municipal and habitat queries once the hardest comparison cases are routed elsewhere.
  • The performance gap between real-API tool use and synthetic benchmarks suggests that future agent training should prioritize production-style error handling and multi-turn recovery.

Load-bearing premise

The 93 tasks and sixteen tools accurately represent the data-wrangling and analysis challenges that environmental scientists actually face.

What would settle it

An agent that scores above 50 percent on the close-value comparison subset or that maintains high success when the same tasks are replaced by logs of actual scientist workflows.

Figures

Figures reproduced from arXiv: 2606.12821 by Devika Jain, Diego Prieto-Herr\'aez, Gabriel Diaz-Ireland, Javier Vel\'azquez, Mario Garc\'ia Peces.

Figure 1
Figure 1. Figure 1: GeoNatureAgent Benchmark system architecture. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Scoring pipeline. Each case is evaluated by up to [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: GeoNatureAgent Benchmark v5 leaderboard. Accu [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Binary accuracy vs partial-credit check score. The Cost-Accuracy Trade-off (bubble size = total tokens) [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Cost-accuracy trade-off. Bubble size is propor [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Category difficulty —mean accuracy across all seven [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Token consumption vs. accuracy. Token volume is [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Accuracy by category across the seven models. The [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
read the original abstract

Environmental scientists spend disproportionate effort on data wrangling rather than analysis, and AI agents that automate geospatial workflows remain unvalidated: no benchmark evaluates agents operating through structured tool calling against real APIs. We introduce the GeoNatureAgent Benchmark, the first benchmark for environmental analysis agents that operate via structured tool calls to a production-style geospatial API. It comprises 93 tasks across 18 categories, covering municipality analysis, multi-turn conversation, spatial reasoning, cross-indicator synthesis, error handling and recovery, ranking, comparison, multilingual understanding, habitat analysis, and task rejection. Tasks are evaluated against an open, self-hostable API serving three environmental indicators across Spain and Portugal via sixteen tools. We evaluate seven LLMs (Claude Sonnet 4, DeepSeek V3.2, GLM-5, Gemini 2.5 Pro, Qwen3-235B, GPT-OSS-120B, Llama 4 Scout) under three temperature-1.0 seeds, reporting capability and per-case cost as orthogonal axes. We find: (1) Claude Sonnet 4 leads at 60.8% +/- 0.8%, followed by DeepSeek V3.2 at 56.3% +/- 3.1%, with no other model above 51%; (2) the cost-accuracy Pareto frontier is occupied mostly by open-weight models, with DeepSeek V3.2 offering 93% of Claude's capability at 11x lower cost ($0.011/case); (3) comparison tasks remain universally unsolved (0% on close-value comparisons), exposing systematic reasoning limits; and (4) structured tool calling against a real API is more discriminative than general-purpose GIS benchmarks, with accuracies 25-35 points lower. We further show extensibility by integrating BigEarthNet V2 land cover for Portugal alongside Spanish CO2 and erosion indicators. The benchmark, harness, and self-hostable API are publicly available.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces the GeoNatureAgent Benchmark, the first benchmark for LLM agents performing environmental geospatial analysis via structured tool calls to a production-style API serving three indicators across Spain and Portugal. It comprises 93 tasks across 18 categories (municipality analysis, multi-turn conversation, spatial reasoning, cross-indicator synthesis, error handling, ranking, comparison, multilingual understanding, habitat analysis, task rejection) and evaluates seven models (Claude Sonnet 4, DeepSeek V3.2, GLM-5, Gemini 2.5 Pro, Qwen3-235B, GPT-OSS-120B, Llama 4 Scout) over three temperature-1.0 seeds. Key findings include Claude Sonnet 4 at 60.8% ± 0.8% success, DeepSeek V3.2 at 56.3% ± 3.1% (11x lower cost), no other model above 51%, universal 0% on close-value comparisons, and the claim that the benchmark is 25-35 points more discriminative than general GIS benchmarks. The benchmark, harness, and self-hostable API are released publicly, with an extensibility demonstration using BigEarthNet V2.

Significance. If the task set and API faithfully represent real environmental workflows, the work supplies a reproducible, domain-specific benchmark that isolates systematic LLM limitations (especially in comparison reasoning) and demonstrates cost-accuracy trade-offs favoring open-weight models. The public release of the benchmark, harness, and API is a concrete strength that enables community follow-up and extensibility experiments.

major comments (2)
  1. [Benchmark construction] Benchmark construction section: No description is given of how the 93 tasks were selected, whether they underwent expert review by environmental scientists, or how their distribution (e.g., municipality analysis, error recovery) was validated against published case studies or logged workflows; this directly undermines the central claim that the benchmark 'faithfully capture[s] the data-wrangling and analysis challenges environmental scientists actually face' and that it is more discriminative than general GIS benchmarks.
  2. [Evaluation methodology] Evaluation methodology: The manuscript supplies no details on ground-truthing of API responses, precise success criteria for multi-turn or error-recovery tasks, or controls for selection bias in the 93-task set; without these, the headline percentages (Claude Sonnet 4 at 60.8% ± 0.8%, 0% on close-value comparisons) cannot be confidently interpreted as measures of agent capability.
minor comments (2)
  1. [Results] Clarify whether the three seeds were run for every model or only a subset, and report per-model variance consistently.
  2. [Discussion] The abstract states 'structured tool calling against a real API is more discriminative... with accuracies 25-35 points lower'; ensure the main text supplies the exact baseline benchmark, models, and task overlap used for this comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on benchmark construction and evaluation methodology. We agree that the manuscript would benefit from expanded details in these areas to better support our claims of faithful representation of environmental workflows and reliable performance metrics. We will revise accordingly.

read point-by-point responses

  1. Referee: [Benchmark construction] Benchmark construction section: No description is given of how the 93 tasks were selected, whether they underwent expert review by environmental scientists, or how their distribution (e.g., municipality analysis, error recovery) was validated against published case studies or logged workflows; this directly undermines the central claim that the benchmark 'faithfully capture[s] the data-wrangling and analysis challenges environmental scientists actually face' and that it is more discriminative than general GIS benchmarks.

    Authors: We acknowledge the omission in the current manuscript. In the revision we will add a new subsection under Benchmark Construction that describes the task selection process: the 93 tasks were derived from representative environmental geospatial workflows documented in peer-reviewed literature on ecological indicator analysis and remote sensing applications across the Iberian Peninsula; the author team (which includes environmental scientists) performed internal expert review for ecological plausibility and category balance; and category distribution was cross-checked against common analysis patterns in published case studies. This addition will directly support the claim of faithful capture and explain the observed discriminativeness relative to general GIS benchmarks. revision: yes

  2. Referee: [Evaluation methodology] Evaluation methodology: The manuscript supplies no details on ground-truthing of API responses, precise success criteria for multi-turn or error-recovery tasks, or controls for selection bias in the 93-task set; without these, the headline percentages (Claude Sonnet 4 at 60.8% ± 0.8%, 0% on close-value comparisons) cannot be confidently interpreted as measures of agent capability.

    Authors: We agree these details are required for confident interpretation. The revised manuscript will expand the Evaluation Methodology section to specify: ground-truth responses were generated by executing each task against the self-hostable API with manually verified correct outputs; success criteria are defined per category (e.g., for multi-turn tasks, all required tool calls must be issued correctly and the final synthesized answer must match the ground truth within tolerance; for error-recovery, the agent must issue a valid corrective tool call without fabricating data); and selection bias was mitigated by exhaustive category coverage with uniform sampling within categories. These clarifications will enable readers to interpret the reported accuracies, including the 0% on close-value comparisons, as reliable measures of agent capability. revision: yes

Circularity Check

0 steps flagged

No circularity: results are direct empirical measurements on fixed external tasks and API

full rationale

The paper defines a benchmark of 93 tasks and 16 tools against a production-style API, then reports measured success rates (e.g., Claude Sonnet 4 at 60.8%) for seven LLMs. No equations, fitted parameters, predictions, or derivations exist that could reduce to inputs by construction. No self-citations are invoked as load-bearing premises for uniqueness or ansatzes. All claims rest on direct evaluation against the stated task set and API, making the work self-contained and non-circular by the enumerated criteria.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The work rests on standard AI benchmarking practices and the authors' construction of new tasks and an API; no free parameters are fitted, no new physical or mathematical entities are postulated, and the central claims are empirical measurements rather than derivations.

axioms (1)
  • domain assumption Success rate on a fixed task set with multiple temperature-1.0 seeds provides a stable measure of agent capability.
    Invoked to report means and standard deviations across three runs.

pith-pipeline@v0.9.1-grok · 5920 in / 1370 out tokens · 37216 ms · 2026-06-27T07:16:27.545438+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 23 canonical work pages · 2 internal anchors

  1. [1]

    Semantic enhancement and change consistency network for semantic change detection in remote sensing images,

    Temitope Akinboyewa, Zhenlong Li, Huan Ning, and M. Naser Lessani. 2025. GIS Copilot: Towards an Autonomous GIS Agent for Spatial Analysis. International Journal of Digital Earth 18, 1 (2025), 2497489. doi:10.1080/17538947.2025.2497489 2https://github.com/gabrielireland/GeoNatureAgent_Benchmark Preprint Diaz-Ireland et al

    work page doi:10.1080/17538947.2025.2497489 2025

  2. [2]

    Feenstra, Conner Arnold, Jan DeWitt, Natalie C

    Christian Michael Arnold, Andrew Alini, Jonathan Wang, Pieter M. Feenstra, Conner Arnold, Jan DeWitt, Natalie C. Ritsema, Jung Hyun Yae, Gary C. Bor- chardt, Boris Katz, Andrei Barbu, and Brian Cheung. 2026. MapQA: A Map- Question-Answering Benchmark for Visual Language Model Reasoning. https: //openreview.net/forum?id=dOISCbmkmG

    2026

  3. [3]

    Rishi Bommasani, Percy Liang, and Tony Lee. 2023. Holistic Evaluation of Language Models. Annals of the New York Academy of Sciences 1525, 1 (2023), 140–

    2023

  4. [4]

    arXiv:https://nyaspubs.onlinelibrary.wiley.com/doi/pdf/10.1111/nyas.15007 doi:10.1111/nyas.15007

    work page doi:10.1111/nyas.15007

  5. [5]

    Kai Norman Clasen, Leonard Hackel, Tom Burgert, Gencer Sumbul, Begüm Demir, and Volker Markl. 2025. reBen: Refined BigEarthNet Dataset for Remote Sensing Image Analysis. In IGARSS 2025 – 2025 IEEE International Geoscience and Remote Sensing Symposium. 1264–1268. doi:10.1109/IGARSS55030.2025.11242834

    work page doi:10.1109/igarss55030.2025.11242834 2025

  6. [6]

    Gabriel Diaz-Ireland, Diego Prieto-Herráez, Mario García Peces, Javier Velázquez, and Devika Jain. 2026. GeoNatureAgent Benchmark — Dataset (tasks, results, derived environmental indicators). doi:10.5281/zenodo.20450995 Concept DOI; resolves to the latest version

    work page doi:10.5281/zenodo.20450995 2026

  7. [7]

    Gabriel Diaz-Ireland, Diego Prieto-Herráez, Mario García Peces, Javier Velázquez, and Devika Jain. 2026. GeoNatureAgent Benchmark — Evaluation Harness, Agent and Self-Hostable Geospatial API. doi:10.5281/zenodo.20450997 Concept DOI; resolves to the latest version

    work page doi:10.5281/zenodo.20450997 2026

  8. [8]

    Mahir Labib Dihan, Md Tanvir Hassan, Md Tanvir Parvez, Md Hasebul Hasan, Md Almash Alam, Muhammad Aamir Cheema, Mohammed Eunus Ali, and Md Rizwan Parvez. 2025. MapEval: A Map-Based Evaluation of Geo-Spatial Reasoning in Foundation Models. In Proceedings of the 42nd International Confer- ence on Machine Learning (Vancouver, Canada) (ICML’25). JMLR.org, Art...

    2025

  9. [9]

    Google Research. 2025. Google Earth AI: Unlocking Geospatial Insights with Foundation Models and Cross-Modal Reasoning. Google Research Blog. https://research.google/blog/google-earth-ai-unlocking-geospatial- insights-with-foundation-models-and-cross-modal-reasoning/

    2025

  10. [10]

    Yu Huang, Liang Guo, Wanqian Guo, Zhe Tao, Yang Lv, Zhihao Sun, and Dongfang Zhao. 2024. EnviroExam: Benchmarking Environmental Science Knowledge of Large Language Models. arXiv:2405.11265 [cs.CL] doi:10.48550/arXiv.2405.11265

    work page doi:10.48550/arxiv.2405.11265 2024

  11. [11]

    Chia Hsiang Kao, Wenting Zhao, Shreelekha Revankar, Samuel Speas, Snehal Bhagat, Rajeev Datta, Cheng Perng Phoo, Utkarsh Mall, Carl Vondrick, Kavita Bala, and Bharath Hariharan. 2025. Towards LLM Agents for Earth Observation. arXiv:2504.12110 [cs.AI] doi:10.48550/arXiv.2504.12110

    work page doi:10.48550/arxiv.2504.12110 2025

  12. [12]

    Kingdom of Spain. 2025. Real Decreto 214/2025 on the Regulation of the Carbon Footprint Register. Boletín Oficial del Estado. https://www.boe.es/buscar/doc. php?id=BOE-A-2025-7439

    2025

  13. [13]

    Varvara Krechetova and Denis Kochedykov. 2025. GeoBenchX: Benchmarking LLMs in Agent Solving Multistep Geospatial Tasks. In Proceedings of the 1st ACM SIGSPATIAL International Workshop on Generative and Agentic AI for Multi- Modality Space-Time Intelligence (GeoGenAgent ’25). Association for Computing Machinery, New York, NY, USA, 27–35. doi:10.1145/37649...

    work page doi:10.1145/3764915.3770721 2025

  14. [14]

    Alexandre Lacoste, Nils Lehmann, Pau Rodriguez, Evan David Sherwin, Hannah Kerner, Björn Lütjens, Jeremy Irvin, David Dao, Hamed Alemohammad, Alexan- dre Drouin, Mehmet Gunturkun, Gabriel Huang, David Vazquez, Dava Newman, Yoshua Bengio, Stefano Ermon, and Xiao Xiang Zhu. 2023. GEO-Bench: Toward Foundation Models for Earth Monitoring. In Proceedings of th...

    2023

  15. [15]

    Chaehong Lee, Varatheepan Paramanayakam, Andreas Karatzas, Yanan Jian, Michael Fore, Heming Liao, Fuxun Yu, Ruopu Li, Iraklis Anagnostopoulos, and Dimitrios Stamoulis. 2025. Multi-Agent Geospatial Copilots for Remote Sensing Workflows. In IGARSS 2025 – 2025 IEEE International Geoscience and Remote Sensing Symposium. 1084–1089. doi:10.1109/IGARSS55030.2025...

    work page doi:10.1109/igarss55030.2025.11243915 2025

  16. [16]

    Zhenlong Li and Huan Ning. 2023. Autonomous GIS: The Next-Generation AI-Powered GIS. International Journal of Digital Earth 16, 2 (2023), 4668–4686. doi:10.1080/17538947.2023.2278895

    work page doi:10.1080/17538947.2023.2278895 2023

  17. [17]

    Qianqian Luo, Qingming Lin, Liuchang Xu, Sensen Wu, Ruichen Mao, Chao Wang, Hailin Feng, Bo Huang, and Zhenhong Du. 2026. GeoJSON Agents: A Multi-Agent LLM Architecture for Geospatial Analysis—Function Calling vs. Code Generation. Big Earth Data 0, 0 (2026), 1–55. doi:10.1080/20964471.2026.2615511

    work page doi:10.1080/20964471.2026.2615511 2026

  18. [18]

    Valerio Marsocci, Yuru Jia, Georges Le Bellier, David Kerekes, Liang Zeng, Se- bastian Hafner, Sebastian Gerard, Eric Brune, Ritu Yadav, Ali Shibli, Heng Fang, Yifang Ban, Maarten Vergauwen, Nicolas Audebert, and Andrea Nascetti. 2025. PANGAEA: A Global and Inclusive Benchmark for Geospatial Foundation Models. arXiv:2412.04204 [cs.CV] doi:10.48550/arXiv.2...

    work page doi:10.48550/arxiv.2412.04204 2025

  19. [19]

    Grégoire Mialon, Clémentine Fourrier, Thomas Wolf, Yann LeCun, and Thomas Scialom. 2024. GAIA: a benchmark for General AI Assistants. In The Twelfth International Conference on Learning Representations . https://openreview.net/ forum?id=fibxvahvs3

    2024

  20. [20]

    Yu-Erh Pan and Ayesha Siddika Nipu. 2025. LLM-Enhanced Air Quality Monitor- ing Interface via Model Context Protocol. In 2025 7th International Symposium on Advanced Electrical and Communication Technologies (ISAECT) . IEEE, 1–6. doi:10.1109/ISAECT68904.2025.11318775

    work page doi:10.1109/isaect68904.2025.11318775 2025

  21. [21]

    Planet Labs and Anthropic. 2025. Planet and Anthropic Partner to Use Claude’s Advanced AI Capabilities to Turn Geospatial Satellite Imagery into Actionable Insights. BusinessWire. https://www.businesswire.com/news/home/ 20250306606139/en/Planet-and-Anthropic-Partner-to-Use-Claudes-Advanced- AI-Capabilities-to-Turn-Geospatial-Satellite-Imagery-into-Actiona...

    2025

  22. [22]

    Akashah Shabbir, Muhammad Akhtar Munir, Akshay Dudhane, Muham- mad Umer Sheikh, Muhammad Haris Khan, Paolo Fraccaro, Juan Bernabe Moreno, Fahad Shahbaz Khan, and Salman Khan. 2026. ThinkGeo: Evaluat- ing Tool-Augmented Agents for Remote Sensing Tasks. arXiv:2505.23752 [cs.CV] doi:10.48550/arXiv.2505.23752

    work page doi:10.48550/arxiv.2505.23752 2026

  23. [23]

    Naomi Simumba, Nils Lehmann, Paolo Fraccaro, Hamed Alemohammad, Geeth De Mel, Salman Khan, Manil Maskey, Nicolas Longepe, Xiao Xiang Zhu, Hannah Kerner, Juan Bernabe-Moreno, and Alexandre Lacoste. 2026. GEO- Bench-2: From Performance to Capability, Rethinking Evaluation in Geospatial AI. arXiv:2511.15658 [cs.CV] doi:10.48550/arXiv.2511.15658

    work page doi:10.48550/arxiv.2511.15658 2026

  24. [24]

    Dimitrios Stamoulis and Diana Marculescu. 2025. Geo-OLM: Enabling Sustainable Earth Observation Studies with Cost-Efficient Open Language Models and State- Driven Workflows. InProceedings of the 2025 ACM SIGCAS/SIGCHI Conference on Computing and Sustainable Societies (COMPASS ’25). Association for Computing Machinery, New York, NY, USA, 608–619. doi:10.11...

    work page doi:10.1145/3715335.3735494 2025

  25. [25]

    Daniela Szwarcman, Sujit Roy, Paolo Fraccaro, Thorsteinn Elí Gíslason, Benedikt Blumenstiel, Rinki Ghosal, Pedro Henrique de Oliveira, Joao Lucas de Sousa Almeida, Rocco Sedona, Yanghui Kang, Srija Chakraborty, Sizhe Wang, Carlos Gomes, Ankur Kumar, Vishal Gaur, Myscon Truong, Denys Godwin, Sam Khallaghi, Hyunho Lee, Chia-Yu Hsu, Ata Akbari Asanjan, Besar...

    work page doi:10.1109/tgrs.2025.3642610 2026

  26. [26]

    Thinh Hung Truong, Jey Han Lau, and Jianzhong Qi. 2026. GPSBench: Do Large Language Models Understand GPS Coordinates? arXiv:2602.16105 [cs.AI] doi:10.48550/arXiv.2602.16105

    work page doi:10.48550/arxiv.2602.16105 2026

  27. [27]

    Iqramul Hoque, Shahriyar Zaman Ridoy, Mo- hammed Eunus Ali, Majd Hawasly, Mohammad Raza, and Md Rizwan Parvez

    Azmine Toushik Wasi, Wahid Faisal, Abdur Rahman, Mahfuz Ahmed Anik, Munem Shahriar, Mohsin Mahmud Topu, Sadia Tasnim Meem, Rahatun Nesa Priti, Sabrina Afroz Mitu, Md. Iqramul Hoque, Shahriyar Zaman Ridoy, Mo- hammed Eunus Ali, Majd Hawasly, Mohammad Raza, and Md Rizwan Parvez

  28. [28]

    SpatiaLab: Can Vision-Language Models Perform Spatial Reasoning in the Wild? arXiv:2602.03916 [cs.CV] doi:10.48550/arXiv.2602.03916

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2602.03916

  29. [29]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing Reasoning and Acting in Language Mod- els. In Proceedings of the 11th International Conference on Learning Representations, ICLR 2023. doi:10.48550/arXiv.2210.03629

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2210.03629 2023

  30. [30]

    Yifan Zhang, Jingxuan Li, Zhiyun Wang, Zhengting He, Qingfeng Guan, Jianfeng Lin, and Wenhao Yu. 2025. Geospatial Large Language Model Trained with a Simulated Environment for Generating Tool-Use Chains Autonomously. Interna- tional Journal of Applied Earth Observation and Geoinformation 136 (2025), 104312. doi:10.1016/j.jag.2024.104312

    work page doi:10.1016/j.jag.2024.104312 2025

  31. [31]

    Yuanxin Zhang, Sijie Lin, Yaxin Xiong, Nan Li, Lijin Zhong, Longzhen Ding, and Qing Hu. 2025. Fine-Tuning Large Language Models for Interdisciplinary Environmental Challenges. Environmental Science and Ecotechnology 27 (2025), 100608. doi:10.1016/j.ese.2025.100608

    work page doi:10.1016/j.ese.2025.100608 2025

  32. [32]

    Yifan Zhang, Cheng Wei, Zhengting He, and Wenhao Yu. 2024. GeoGPT: An Assistant for Understanding and Processing Geospatial Tasks. International Journal of Applied Earth Observation and Geoinformation 131 (2024), 103976. doi:10.1016/j.jag.2024.103976 A Online Resources The GeoNatureAgent Benchmark code, evaluation harness, and self-hostable API are availa...

    work page doi:10.1016/j.jag.2024.103976 2024