Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Codified expert workflows let an AI agent automate remote sensing analysis with over 95% task success.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 15:50 UTC pith:6GH2K6NO

load-bearing objection The integration is sensible, but the evaluation leaks the answers: the agent retrieves expert templates for the exact tasks it's graded on, and the ThinkGeo adaptation uses the test set's gold traces. the 3 major comments →

arxiv 2512.15231 v3 pith:6GH2K6NO submitted 2025-12-17 cs.AI

CangLing-KnowFlow: A Unified Knowledge-and-Flow-fused Agent for Comprehensive Remote Sensing Applications

Zhengchao Chen , Haoran Wang , Jing Yao , Jianshe Zhang , Pedram Ghamisi , Jun Zhou , Peter M. Atkinson , Bing Zhang This is my paper
classification cs.AI
keywords remote sensingAI agentprocedural knowledge baseLLM planningworkflow orchestrationdynamic adjustmentevolutionary memoryEarth observation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper is trying to establish that the bottleneck for AI agents in remote sensing is not raw language-model intelligence but the absence of codified, expert-validated procedural knowledge. It proposes an agent that retrieves a workflow template from a knowledge base of 1,008 expert workflows across 162 tasks, executes it as a directed graph of tool calls, and repairs failures by inserting, replacing, or re-parameterizing steps, with an experience memory that turns successes and failures into reusable knowledge. On the authors' new benchmark of 324 gold workflows, the framework, powered by a strong commercial LLM, reaches 96.1% and 95.1% task success on simple and complex tasks, outpacing free-form reasoning baselines and cutting tool calls by about 38%. If true, this would mean that structured procedural knowledge, not a bigger model, is the decisive factor for reliable scientific automation, and that agents can evolve from accumulated experience.

Core claim

The central claim is that fusing an expert-informed Procedural Knowledge Base with dynamic workflow adjustment and evolutionary memory produces an agent that can plan, execute, and repair long-horizon remote sensing workflows more reliably and efficiently than general-purpose reasoning agents. The authors show that when the agent starts from a retrieved workflow template rather than generating a plan from scratch, initial plans are more often correct (First-Pass Accuracy 79.2% on simple tasks), and when execution fails, the agent can autonomously diagnose and repair the workflow graph through node replacement, insertion, or parameter modification. The evolutionary memory then solidifies succ

What carries the argument

The central object is the Procedural Knowledge Base, a library of 1,008 expert-validated workflow templates, each represented as a Directed Acyclic Graph whose vertices are tool invocations and whose edges encode dependency constraints chosen by semantic compatibility. A formal planning-language-style tool schema gives every tool preconditions and effects, allowing the agent to check logical consistency, detect failure points, and manipulate the graph. The Dynamic Workflow Adjustment module applies three graph operations—node replacement, node insertion, and parameter modification—to repair failed workflows, and the Evolutionary Memory Module stores execution traces and generalizes them into

Load-bearing premise

The load-bearing premise is that the benchmark is a fair out-of-sample test, because its ground-truth workflows come from the same 162 tasks used to populate the agent's knowledge base; if it measures template retrieval rather than genuine planning, the headline success rate overstates generality.

What would settle it

Hold out a random subset of the 162 task types: remove their templates from the knowledge base before evaluation, then run the agent on those held-out tasks. If the task success rate falls to within a few points of a free-form reasoning baseline, the framework's gains are retrieval rather than adaptive planning; if it stays high, the agent genuinely generalizes.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Remote sensing analyses from data acquisition to change detection and statistical reporting can run end-to-end without human intervention at success rates above 90%.
  • Planning hallucinations—plans that look plausible but are scientifically unsound—should largely disappear when the agent is anchored to expert-validated workflows.
  • Organizations can use weaker or cheaper LLMs and still outperform stronger models used in a free-form manner, because the knowledge base carries the domain expertise.
  • Runtime failures such as cloudy data or tool incompatibility can be autonomously repaired, and the repair strategies improve with each task execution.
  • The benchmark provides a reusable protocol for measuring workflow-level planning quality, not just final answers.

Where Pith is reading between the lines

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

  • Because the benchmark and knowledge base are built from the same 162 tasks, the reported success may partly reflect retrieval of near-answer templates rather than emergent planning; a held-out task evaluation would separate these.
  • The adaptation to an external benchmark extracts workflow logic from the benchmark's own annotations, which could leak ground-truth steps; if so, the transferability results are optimistic.
  • A longitudinal deployment should show compounding gains from the evolutionary memory—higher first-pass accuracy and fewer tool calls over thousands of tasks—which a single-session evaluation cannot capture.
  • The same architecture is probably transferable to other tool-intensive scientific domains wherever expert procedures can be formalized as graph templates.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes CangLing-KnowFlow, an LLM-agent framework for remote sensing that combines a Procedural Knowledge Base (PKB) of 1,008 expert-validated workflow templates over 162 tasks, dynamic workflow adjustment, and an evolutionary memory module. The authors evaluate the framework on a self-constructed benchmark, KnowFlow-Bench, and on the existing ThinkGeo benchmark, reporting large Task Success Rate improvements over ReAct and Reflexion baselines, with peak TSRs of 96.1% (simple) and 95.1% (complex) using Claude 4 Sonnet. The core problem is that the evaluation is not out-of-sample: KnowFlow-Bench is built from the very same 162 tasks used to populate the PKB, and the ThinkGeo adaptation is configured using the benchmark's ReAct-style gold annotations. The reported gains therefore cannot be interpreted as evidence of planning capability.

Significance. If the empirical claims were valid, the paper would be a useful contribution to domain-specific LLM agents: the PKB is a substantial curated resource, the multi-backbone evaluation is broad, and the framework combines retrieval, repair, and memory in a coherent architecture. However, the central evidence is undermined by benchmark construction. The TSR advantage over ReAct/Reflexion largely measures whether the agent can retrieve and instantiate an expert template for a task that is guaranteed to have one in the PKB; it does not measure whether the agent can plan under uncertainty. The ThinkGeo result is similarly contaminated. The resource may still be valuable, but the manuscript's headline claims are not supported by the reported experiments.

major comments (3)
  1. [§4.1.2, §4.2.1, Algorithm 1] KnowFlow-Bench is constructed from the same 162 task descriptions used to populate the PKB: §4.1.2 states that the 1,008 templates are curated from 162 LLM-generated RS task descriptions, and §4.2.1 states that the benchmark is 'built upon the 162 tasks curated from our PKB'. Algorithm 1 retrieves a workflow template from the PKB for every incoming task before execution. Since each of the 162 tasks has 2–10 expert templates, every benchmark instance has a near-answer template available at retrieval time. The TSR gains in Table 1 therefore measure template retrieval and instantiation fidelity, not emergent planning or tool-orchestration capability. ReAct and Reflexion have no equivalent access, so the comparison is structurally unfair and the headline 'at least 4% TSR improvement' is uninterpretable as evidence of agent capability.
  2. [§4.2.2, Table 2] The ThinkGeo experiment is not an out-of-sample generalization test. The authors state that adapting CangLing-KnowFlow to ThinkGeo 'entailed extracting the procedural workflow logic from its ReAct-style annotations and restricting the agent's action space strictly to the 14 tools'. This injects the benchmark's gold reasoning traces into the agent before evaluation. No held-out split is reported, so Table 2's 20.3% TSR (vs. 9.0% for ReAct) can reflect test-set information leakage rather than transferable capability. The Section 5.2 claim that ThinkGeo is an 'unseen case' is contradicted by the paper's own §4.2.2 description.
  3. [§5.2, Table 1] No error bars, confidence intervals, or significance tests are reported for any metric. The performance differences between the proposed method and Reflexion can be small in absolute terms (e.g., 3.5 percentage points on simple tasks for Claude 4 Sonnet; 4.2–4.3 points on complex tasks for several backbones). Without information about the number of repeated runs or per-task variance, the claim of 'significant improvement' is not statistically established even aside from the leakage issues above. The paper should report the number of independent trials and appropriate uncertainty measures.
minor comments (5)
  1. [Throughout] There are several typos and grammatical issues: 'blancing' (Section 1), 'generalizaion' (Related Work), 'effectivity' and 'occured' (Section 5), 'irrelevant' in the experimental setup. The paper would benefit from a careful proofread.
  2. [Table 2] Some NTC/NI entries are malformed or approximate (e.g., '1250', '∼1350', '∼9.5'). If these are ranges or estimates, the notation should be defined; otherwise the table should give exact values.
  3. [§5.5] The qualitative case study is presented as illustrative, but it reads as evidence for the framework's superiority. It should be explicitly labeled as a non-quantitative illustration, or better, replaced with a systematic error analysis across the full benchmark.
  4. [Code Website] The 'Code Website' line after the abstract does not contain a URL. If code is being released, a working link should be included; if not, the line should be removed.
  5. [References] Several model names in Table 1 (e.g., Claude 4 Sonnet, GPT-5) do not have corresponding citations in the reference list. Please add proper citations or clearly mark them as unreleased models.

Circularity Check

2 steps flagged

Not out-of-sample: KnowFlow-Bench uses the same 162 tasks as the PKB the agent retrieves from, and §4.2.2 configures the agent from ThinkGeo's ReAct-style annotations, so TSR gains partly measure answer retrieval.

specific steps
  1. self definitional [§4.1.2, §4.2.1, Algorithm 1 line 2]
    "KnowFlow-Bench is built upon the 162 tasks curated from our PKB. | The process began by generating a diverse set of 162 realistic RS task descriptions ... For each task, domain experts then manually annotated between two and 10 plausible, high-level solution workflows."

    The same 162 task descriptions define both the PKB (2–10 expert templates per task, §4.1.2) and KnowFlow-Bench (2 gold workflows per task, §4.2.1). At runtime Algorithm 1 executes 'W_template ← RetrieveWorkflowTemplate(Guser, K_proc)', so for every benchmark task the agent retrieves an expert-written high-level plan for that exact task. The reported TSR therefore measures retrieval and template instantiation, not planning from first principles. ReAct and Reflexion have no such access, making the headline comparison (e.g., 95.1% vs 90.6%) uninterpretable as evidence of superior planning capability.

  2. other [§4.2.2, Table 2]
    "This adaptation entailed extracting the procedural workflow logic from its ReAct-style annotations and restricting the agent’s action space strictly to the 14 tools provided by the benchmark."

    The paper describes ThinkGeo as evaluating 'the ability to follow a ReAct-style reasoning and execution chain'; its annotations are the benchmark's ground-truth traces. If the agent is configured by extracting procedural workflow logic from those traces, the test answers themselves (or their structure) are used to build the agent. No held-out split is reported, so Table 2's 'unseen environment' claim (TSR 20.3% vs ReAct 9.0%) is not a clean generalization result: the configuration step may inject the gold-trace structure directly into the agent.

full rationale

The circularity is not in the framework's internal derivation (MDP, DAG, PDDL schemas are self-contained formalizations), and there is no load-bearing self-citation chain. The problem is the evaluation design. KnowFlow-Bench is "built upon the 162 tasks curated from our PKB", and Algorithm 1 retrieves a PKB template for the user's task, so every benchmark task has a near-answer template in the agent's knowledge base. The 4%+ margin over Reflexion is therefore at least partly retrieval fidelity. The ThinkGeo adaptation admits extracting workflow logic from the benchmark's ReAct-style annotations, which, without a held-out split, leaks the test structure. These two issues affect the paper's central empirical claim. The score is not 10 because the PKB templates are abstract (high-level steps) while KnowFlow-Bench gold workflows are concrete function calls with parameters, so the agent still performs real instantiation and execution work; and the ThinkGeo sentence is ambiguous about whether the extracted 'logic' includes gold solutions. Still, the load-bearing results are not out-of-sample, so the empirical claims reduce substantially to the paper's own knowledge-base construction choices.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

No formal proof is offered. The framework's effectiveness rests on assumptions that the PKB is expert-correct, representative, and non-overlapping with the evaluation. The last assumption is violated by construction, and the phi_LLM edge function is unvalidated. No new physical entities are introduced.

free parameters (1)
  • Composition and overlap of the workflow library = 162 tasks, 1,008 templates (2–10 per task)
    The size and task coverage of the PKB are hand-chosen and share the same task pool as KnowFlow-Bench; this design choice largely determines the reported success rates and is not fit to independent external data.
axioms (5)
  • domain assumption PKB workflow templates are expert-validated and scientifically correct.
    The paper relies on the expert annotation of 1,008 templates as ground truth for scientific validity; there is no independent audit, inter-annotator agreement, or release of the templates.
  • domain assumption The 162 task descriptions generated by Gemini 2.5 Pro are realistic and representative of practical remote-sensing tasks.
    The entire PKB and benchmark are anchored to LLM-generated task descriptions (§4.1.2); representativeness is asserted, not demonstrated.
  • ad hoc to paper Tool preconditions/effects can be faithfully captured in a PDDL-like schema, and the LLM function phi_LLM correctly infers edge compatibility.
    Equation (2) defines edges based on an LLM boolean indicator; no accuracy, reliability, or failure analysis is provided for this component.
  • ad hoc to paper Topical retrieval of the most similar workflow template yields a valid plan for the current task.
    Algorithm 1 assumes the retrieved template is correct enough to serve as the primary procedural guide; retrieval accuracy is not separately evaluated.
  • domain assumption Success solidification and failure attribution improve future performance without overfitting to the evaluation set.
    The evolutionary memory enriches the same PKB used to build the benchmark, creating a circular improvement loop that is not tested on genuinely new task types.

pith-pipeline@v1.3.0-alltime-deepseek · 21120 in / 12558 out tokens · 131582 ms · 2026-08-03T15:50:26.760543+00:00 · methodology

0 comments
read the original abstract

The automated and intelligent processing of massive remote sensing (RS) datasets is critical in Earth observation (EO). Existing automated systems are normally task-specific, lacking a unified framework to manage diverse, end-to-end workflows--from data preprocessing to advanced interpretation--across diverse RS applications. To address this gap, this paper introduces CangLing-KnowFlow, a unified intelligent agent framework that integrates a Procedural Knowledge Base (PKB), Dynamic Workflow Adjustment, and an Evolutionary Memory Module. The PKB, comprising 1,008 expert-validated workflow cases across 162 practical RS tasks, guides planning and substantially reduces hallucinations common in general-purpose agents. During runtime failures, the Dynamic Workflow Adjustment autonomously diagnoses and replans recovery strategies, while the Evolutionary Memory Module continuously learns from these events, iteratively enhancing the agent's knowledge and performance. This synergy enables CangLing-KnowFlow to adapt, learn, and operate reliably across diverse, complex tasks. We evaluated CangLing-KnowFlow on the KnowFlow-Bench, a novel benchmark of 324 workflows inspired by real-world applications, testing its performance across 13 top Large Language Model (LLM) backbones, from open-source to commercial. Across all complex tasks, CangLing-KnowFlow surpassed the Reflexion baseline by at least 4% in Task Success Rate. As the first most comprehensive validation along this emerging field, this research demonstrates the great potential of CangLing-KnowFlow as a robust, efficient, and scalable automated solution for complex EO challenges by leveraging expert knowledge (Knowledge) into adaptive and verifiable procedures (Flow).

Figures

Figures reproduced from arXiv: 2512.15231 by Bing Zhang, Haoran Wang, Jianshe Zhang, Jing Yao, Jun Zhou, Pedram Ghamisi, Peter M. Atkinson, Zhengchao Chen.

Figure 1
Figure 1. Figure 1: Conceptual comparison of different agent architectures for remote sensing. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overall architecture of the CangLing-KnowFlow framework, illustrating the inter [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Structure of the Procedural Knowledge Base. (a) An illustration of a hierarchical workflow [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Schematic of the Dynamic Workflow Adjustment and Evolutionary Memory. (a) The [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: CangLing-KnowFlow provides comprehensive support for 162 diverse remote sensing [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparing the results in the visualization, ReAct missed Step 3 and obtained a poor [PITH_FULL_IMAGE:figures/full_fig_p023_6.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. RemoteAgent: Bridging Vague Human Intents and Earth Observation with RL-based Agentic MLLMs

    cs.CV 2026-04 unverdicted novelty 7.0

    RemoteAgent uses RL fine-tuning on VagueEO to align MLLMs for vague EO intent recognition, handling simple tasks internally and routing dense predictions to tools via Model Context Protocol.

Reference graph

Works this paper leans on

45 extracted references · 17 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Deep learn- ing in remote sensing image fusion: Methods, protocols, data, and future per- spectives.IEEE Geoscience and Remote Sensing Magazine, 13:269–310, 2024

    GemineVivone, Liang-JianDeng, ShangqiDeng, DanfengHong, MenghuiJiang, Chenyu Li, Wei Li, Huanfeng Shen, Xiao Wu, Jin-Liang Xiao, et al. Deep learn- ing in remote sensing image fusion: Methods, protocols, data, and future per- spectives.IEEE Geoscience and Remote Sensing Magazine, 13:269–310, 2024

    2024

  2. [2]

    Progress and challenges in intelligent remote sensing satellitesystems.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 15:1814–1822, 2022

    Bing Zhang, Yuanfeng Wu, Boya Zhao, Jocelyn Chanussot, Danfeng Hong, Jing Yao, and Lianru Gao. Progress and challenges in intelligent remote sensing satellitesystems.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 15:1814–1822, 2022

    2022

  3. [3]

    Remotely sensed big data: Evolution in model development for information extraction [point of view].Proceedings of the IEEE, 107:2294–2301, 2019

    Bing Zhang, Zhengchao Chen, Dailiang Peng, Jon Atli Benediktsson, Bo Liu, Lei Zou, Jun Li, and Antonio Plaza. Remotely sensed big data: Evolution in model development for information extraction [point of view].Proceedings of the IEEE, 107:2294–2301, 2019

    2019

  4. [4]

    The core concepts and fundamental issues of remote sensing science.National Remote Sensing Bulletin, 29:1–48, 2024

    Bing Zhang, Qinhuo Liu, Xiaoming Li, Liangyun Liu, Bisheng Yang, Letu Husi, Lianru Gao, Wenjuan Zhang, Hao Zhang, Zunjian Bian, Mengjia Qi, Chi Chen, and Huazhe Shang. The core concepts and fundamental issues of remote sensing science.National Remote Sensing Bulletin, 29:1–48, 2024

    2024

  5. [5]

    SegEarth-R1: Geospatial pixel reasoning via large language model.arXiv preprint arXiv:2504.09644, 2025

    Kaiyu Li, Zepeng Xin, Li Pang, Chao Pang, Yupeng Deng, Jing Yao, Guisong Xia, Deyu Meng, Zhi Wang, and Xiangyong Cao. SegEarth-R1: Geospatial pixel reasoning via large language model.arXiv preprint arXiv:2504.09644, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    KnowAgent: Knowledge- augmented planning for LLM-based agents

    Yuqi Zhu, Shuofei Qiao, Yixin Ou, Shumin Deng, Shiwei Lyu, Yue Shen, Lei Liang, Jinjie Gu, Huajun Chen, and Ningyu Zhang. KnowAgent: Knowledge- augmented planning for LLM-based agents. InFindings of the Association for Computational Linguistics: NAACL 2025, pages 3709–3732, 2025

    2025

  7. [7]

    FlowAgent: Achieving compliance and flexibility for workflow agents.arXiv preprint arXiv:2502.14345, 2025

    Yuchen Shi, Siqi Cai, Zihan Xu, Yuei Qin, Gang Li, Hang Shao, Jiawei Chen, Deqing Yang, Ke Li, and Xing Sun. FlowAgent: Achieving compliance and flexibility for workflow agents.arXiv preprint arXiv:2502.14345, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Rs-Agent: Automating remote sensing tasks through intelligent agent.arXiv preprint arXiv:2406.07089, 2024

    Wenjia Xu, Zijian Yu, Boyang Mu, Zhiwei Wei, Yuanben Zhang, Guangzuo Li, and Mugen Peng. Rs-Agent: Automating remote sensing tasks through intelligent agent.arXiv preprint arXiv:2406.07089, 2024. 28

    Pith/arXiv arXiv 2024

  9. [9]

    GeoTool-GPT: a trainable method for fa- cilitating large language models to master GIS tools.International Journal of Geographical Information Science, 39:707–731, 2025

    Cheng Wei, Yifan Zhang, Xinru Zhao, Ziyi Zeng, Zhiyun Wang, Jianfeng Lin, Qingfeng Guan, and Wenhao Yu. GeoTool-GPT: a trainable method for fa- cilitating large language models to master GIS tools.International Journal of Geographical Information Science, 39:707–731, 2025

    2025

  10. [10]

    An LLM agent for automatic geospatial data analysis.arXiv preprint arXiv:2410.18792, 2024

    Yuxing Chen, Weijie Wang, Sylvain Lobry, and Camille Kurtz. An LLM agent for automatic geospatial data analysis.arXiv preprint arXiv:2410.18792, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    LLM- Find: An autonomous gis agent framework for geospatial data retrieval.arXiv e-prints, pages arXiv–2407, 2024

    Huan Ning, Zhenlong Li, Temitope Akinboyewa, and M Naser Lessani. LLM- Find: An autonomous gis agent framework for geospatial data retrieval.arXiv e-prints, pages arXiv–2407, 2024

    2024

  12. [12]

    ShapefileGPT: A multi-agent large language model framework for automated shapefile processing.International Journal of Digital Earth, 18(2):2577884, 2025

    Qingming Lin, Rui Hu, Huaxia Li, Sensen Wu, Yadong Li, Kai Fang, Hailin Feng, Zhenhong Du, and Liuchang Xu. ShapefileGPT: A multi-agent large language model framework for automated shapefile processing.International Journal of Digital Earth, 18(2):2577884, 2025

    2025

  13. [13]

    Multi-Agent geospatial copilots for remote sensing workflows.arXiv preprint arXiv:2501.16254, 2025

    Chaehong Lee, Varatheepan Paramanayakam, Andreas Karatzas, Yanan Jian, Michael Fore, Heming Liao, Fuxun Yu, Ruopu Li, Iraklis Anagnostopoulos, and Dimitrios Stamoulis. Multi-Agent geospatial copilots for remote sensing workflows.arXiv preprint arXiv:2501.16254, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    MineAgent: Towards remote-sensing mineral exploration with multimodal large language models.arXiv preprint arXiv:2412.17339, 2024

    Beibei Yu, Tao Shen, Hongbin Na, Ling Chen, and Denqi Li. MineAgent: Towards remote-sensing mineral exploration with multimodal large language models.arXiv preprint arXiv:2412.17339, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    RSGPT: A remote sensing vision language model and benchmark.ISPRS Journal of Photogrammetry and Remote Sensing, 224:272–286, 2025

    Yuan Hu, Jianlong Yuan, Congcong Wen, Xiaonan Lu, Yu Liu, and Xiang Li. RSGPT: A remote sensing vision language model and benchmark.ISPRS Journal of Photogrammetry and Remote Sensing, 224:272–286, 2025

    2025

  16. [16]

    SkyEyeGPT: Unifying remote sensing vision-language tasks via instruction tuning with large language model

    Yang Zhan, Zhitong Xiong, and Yuan Yuan. SkyEyeGPT: Unifying remote sensing vision-language tasks via instruction tuning with large language model. ISPRS Journal of Photogrammetry and Remote Sensing, 221:64–77, 2025

    2025

  17. [17]

    EarthDial: Turning multi-sensory earth observations to interactive dialogues

    Sagar Soni, Akshay Dudhane, Hiyam Debary, Mustansar Fiaz, Muham- mad Akhtar Munir, Muhammad Sohail Danish, Paolo Fraccaro, Campbell D Watson, Levente J Klein, Fahad Shahbaz Khan, et al. EarthDial: Turning multi-sensory earth observations to interactive dialogues. InProceedings of the Computer Vision and Pattern Recognition Conference, pages14303–14313, 2025. 29

    2025

  18. [18]

    Designing domain-specific agents via hierarchical task abstrac- tion mechanism.arXiv preprint arXiv:2511.17198, 2025

    Kaiyu Li, Jiayu Wang, Zhi Wang, Hui Qiao, Weizhan Zhang, Deyu Meng, and Xiangyong Cao. Designing domain-specific agents via hierarchical task abstrac- tion mechanism.arXiv preprint arXiv:2511.17198, 2025

    arXiv 2025

  19. [19]

    Earth-Agent: Un- locking the full landscape of earth observation with agents.arXiv preprint arXiv:2509.23141, 2025

    Peilin Feng, Zhutao Lv, Junyan Ye, Xiaolei Wang, Xinjie Huo, Jinhua Yu, Wanghan Xu, Wenlong Zhang, Lei Bai, Conghui He, et al. Earth-Agent: Un- locking the full landscape of earth observation with agents.arXiv preprint arXiv:2509.23141, 2025

    arXiv 2025

  20. [20]

    GeoFlow: Agentic workflow automation for geospatial tasks.arXiv preprint arXiv:2508.04719, 2025

    Amulya Bhattaram, Justin Chung, Stanley Chung, Ranit Gupta, Janani Ra- mamoorthy, Kartikeya Gullapalli, Diana Marculescu, and Dimitrios Stamoulis. GeoFlow: Agentic workflow automation for geospatial tasks.arXiv preprint arXiv:2508.04719, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Reflexion: Language agents with verbal reinforcement learning

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36:8634–8652, 2023

    2023

  22. [22]

    ReAct: Synergizing reasoning and acting in lan- guagemodels

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing reasoning and acting in lan- guagemodels. InInternational Conference on Learning Representations (ICLR), pages 1–18, 2023

    2023

  23. [23]

    API-Bank: A Comprehensive Bench- mark for Tool-Augmented LLMs

    Minghao Li, Yingxiu Zhao, Bowen Yu, Feifan Song, Hangyu Li, Haiyang Yu, Zhoujun Li, Fei Huang, and Yongbin Li. API-Bank: A Comprehensive Bench- mark for Tool-Augmented LLMs. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 3102–3116, 2023

    2023

  24. [24]

    ToolAlpaca: Generalized tool learning for language models with 3000 simulated cases.arXiv preprint arXiv:2306.05301, 2023

    Qiaoyu Tang, Ziliang Deng, Hongyu Lin, Xianpei Han, Qiao Liang, Boxi Cao, and Le Sun. ToolAlpaca: Generalized tool learning for language models with 3000 simulated cases.arXiv preprint arXiv:2306.05301, 2023

    Pith/arXiv arXiv 2023

  25. [25]

    Language agent tree search unifies reasoning, acting, and planning in language models

    Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu- Xiong Wang. Language agent tree search unifies reasoning, acting, and planning in language models. InProceedings of the 41st International Conference on Machine Learning, pages 62138–62160, 2024

    2024

  26. [26]

    AutoGen: Enabling next-gen LLM applications via multi-agent conversations

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, et al. AutoGen: Enabling next-gen LLM applications via multi-agent conversations. InFirst Conference on Language Modeling, pages 1–15, 2024. 30

    2024

  27. [27]

    AgentBench: Evaluating LLMs as agents

    Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, et al. AgentBench: Evaluating LLMs as agents. InThe Twelfth International Conference on Learning Repre- sentations, 2024

    2024

  28. [28]

    AutoMind: Adaptive knowledgeable agent for automated data science.arXiv preprint arXiv:2506.10974, 2025

    Yixin Ou, Yujie Luo, Jingsheng Zheng, Lanning Wei, Shuofei Qiao, Jin- tian Zhang, Da Zheng, Huajun Chen, and Ningyu Zhang. AutoMind: Adaptive knowledgeable agent for automated data science.arXiv preprint arXiv:2506.10974, 2025

    arXiv 2025

  29. [29]

    Agent workflow memory.arXiv preprint arXiv:2409.07429, 2024

    Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried, and Graham Neubig. Agent workflow memory.arXiv preprint arXiv:2409.07429, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    AFlow: Automating agentic workflow generation

    Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xionghui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, Sirui Hong, Jinlin Wang, et al. AFlow: Automating agentic workflow generation. InThe Thirteenth International Con- ference on Learning Representations, 2025

    2025

  31. [31]

    GeoGPT:Anassistant for understanding and processing geospatial tasks.International Journal of Applied Earth Observation and Geoinformation, 131:103976, 2024

    YifanZhang, ChengWei, ZhengtingHe, andWenhaoYu. GeoGPT:Anassistant for understanding and processing geospatial tasks.International Journal of Applied Earth Observation and Geoinformation, 131:103976, 2024

    2024

  32. [32]

    GIS Copilot: Towards an autonomous gis agent for spatial analysis.International Journal of Digital Earth, 18(1):2497489, 2025

    Temitope Akinboyewa, Zhenlong Li, Huan Ning, and M Naser Lessani. GIS Copilot: Towards an autonomous gis agent for spatial analysis.International Journal of Digital Earth, 18(1):2497489, 2025

    2025

  33. [33]

    GeoSR: Cognitive-agentic framework for probing geospatial knowledge bound- aries via iterative self-refinement.arXiv preprint arXiv:2508.04080, 2025

    Jinfan Tang, Kunming Wu, Ruifeng Gongxie, Yuya He, and Yuankai Wu. GeoSR: Cognitive-agentic framework for probing geospatial knowledge bound- aries via iterative self-refinement.arXiv preprint arXiv:2508.04080, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Accelerating earth science discovery via multi-Agent LLM systems.Frontiers in Artificial Intelligence, 8:1674927, 2025

    Dmitrii Pantiukhin, Boris Shapkin, Antonia Anna Jost, Ivan Kuznetsov, and Nikolay V Koldunov. Accelerating earth science discovery via multi-Agent LLM systems.Frontiers in Artificial Intelligence, 8:1674927, 2025

    2025

  35. [35]

    A RAG-based multi-agent LLM system for natural hazard resilience and adaptation.arXiv preprint arXiv:2504.17200, 2025

    Yangxinyu Xie, Bowen Jiang, Tanwi Mallick, Joshua David Bergerson, John K Hutchison, Duane R Verner, Jordan Branham, M Ross Alexander, Robert B Ross, Yan Feng, et al. A RAG-based multi-agent LLM system for natural hazard resilience and adaptation.arXiv preprint arXiv:2504.17200, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    ThinkGeo: Evaluating tool- augmented agents for remote sensing tasks.arXiv preprint arXiv:2505.23752, 2025

    Akashah Shabbir, Muhammad Akhtar Munir, Akshay Dudhane, Muham- mad Umer Sheikh, Muhammad Haris Khan, Paolo Fraccaro, Juan Bernabe 31 Moreno, Fahad Shahbaz Khan, and Salman Khan. ThinkGeo: Evaluating tool- augmented agents for remote sensing tasks.arXiv preprint arXiv:2505.23752, 2025

    arXiv 2025

  37. [37]

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodal- ity, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodal- ity, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    Can large language models generate geospatial code?Geo-spatial Information Science, pages 1–35, 2025

    Shuyang Hou, Zhangxiao Shen, Jianyuan Liang, Haoyue Jiao, Anqi Zhao, Yax- ian Qing, Dehua Peng, Zhipeng Gui, Xuefeng Guan, Longgang Xiang, et al. Can large language models generate geospatial code?Geo-spatial Information Science, pages 1–35, 2025

    2025

  39. [39]

    GTA: a benchmark for general tool agents.Advances in Neural Information Processing Systems, 37:75749–75790, 2024

    Jize Wang, Zerun Ma, Yining Li, Songyang Zhang, Cailian Chen, Kai Chen, and Xinyi Le. GTA: a benchmark for general tool agents.Advances in Neural Information Processing Systems, 37:75749–75790, 2024

    2024

  40. [40]

    GPT-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Flo- rencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shya- mal Anadkat, et al. GPT-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  41. [41]

    The Llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Ab- hishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schel- ten, Alex Vaughan, et al. The Llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Pith/arXiv arXiv 2024

  42. [42]

    Mistral 7B.arXiv preprint arXiv:2310.06825, 2023

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, De- vendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7B.arXiv preprint arXiv:2310.06825, 2023

    Pith/arXiv arXiv 2023

  43. [43]

    Deepseek- V3 technical report.arXiv preprint arXiv:2412.19437, 2024

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek- V3 technical report.arXiv preprint arXiv:2412.19437, 2024

    Pith/arXiv arXiv 2024

  44. [44]

    Deepseek-R1 incentivizes reasoning in LLMs through reinforcement learning.Nature, 645:633–638, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-R1 incentivizes reasoning in LLMs through reinforcement learning.Nature, 645:633–638, 2025. 32

    2025

  45. [45]

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.arXiv preprint arXiv:2403.05530, 2024

    Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, An- mol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.arXiv preprint arXiv:2403.05530, 2024. 33

    Pith/arXiv arXiv 2024