GeoSkill:Experience-Driven Hierarchical Skill Learning with Collaborative Revision forGeospatialAgents

# GeoSkill: Experience-Driven Hierarchical Skill Learning with Collaborative Revision for Geospatial Agents
Source: [https://arxiv.org/html/2609.13667](https://arxiv.org/html/2609.13667)
###### Abstract

Geospatial agents are increasingly expected to support recurring and evolving analytical tasks rather than execute isolated workflows\. In such settings, effective agents must distill prior execution experience into reusable geospatial procedural knowledge to guide future planning and tool use\. However, existing memory\-augmented paradigms struggle to summarize both long\-horizon tool\-chain orchestration experience and tool\-level invocation constraints in geospatial analysis, while directly relying on LLM self\-reflection to update experience often leads to misattribution and unreliable revisions\. To address these challenges, we propose GeoSkill, an experience\-driven hierarchical skill learning framework for geospatial agents\. GeoSkill comprises two core components: \(i\) a Hierarchical Skill Bank \(HSB\), consisting of a Planning Skill Bank and a Tool Skill Bank, which respectively distill high\-level task\-planning experience and tool usage constraints, enabling structured representation and cross\-task reuse of historical execution experience; and \(ii\) a Collaborative Trace\-driven Skill Revision \(CTSR\) mechanism, where Judge, Critic, and Refiner collaboratively perform error identification, skill\-level defect localization, and targeted modification, preventing misattributed and unreliable revisions from polluting the skill bank\. GeoSkill learns and validates skills from historical executions during development, and freezes the skill bank for retrieval\-only guidance on unseen tasks during deployment\. Extensive experiments on EarthBench and ThinkGeo demonstrate that GeoSkill effectively transforms historical execution experience into reusable hierarchical skills, improving both end\-to\-end task accuracy and tool\-execution reliability in geospatial tasks\.

1Wuhan University

luo\_han@whu\.edu\.cn, xuxian@whu\.edu\.cn, liuyinhe@whu\.edu\.cn, zhongyanfei@whu\.edu\.cn

![Refer to caption](https://arxiv.org/html/2609.13667v1/fig1.png)Figure 1:Comparison of agent paradigms illustrated with an LST \(Land Surface Temperature\) analysis task\.\(a\) Independent Execution: each task runs a standalone ReAct loop from scratch; experience is lost and wrong data/tool selections repeat across tasks\.\(b\) Intra\-task Reflexion: verbal reflections repair the current trial, but the experience is not transferred to future tasks\.\(c\) Inter\-task Memory: experiences are extracted into a flat pool \(e\.g\., “Use LST product”\), yet critical constraints such as spatial resolution and temporal requirements are missing, causing downstream failures\.\(d\) Hierarchical Skill Evolution \(Ours\): a Hierarchical Skill Bank separates planning skills \(LST analysis workflow\) from tool skills \(LST product constraints\), while Causal Trace\-driven Skill Revision evolves them through validated edits, preserving both workflow structure and execution constraints\.## Introduction

Geospatial analysis is moving beyond static prediction toward multi\-step analytical workflows that require coordinated reasoning over data, tools, and geospatial states\([Munir et al\. 2026](https://arxiv.org/html/2609.13667#bib.bib20)\)\. Recent studies\([Shabbir et al\. 2025](https://arxiv.org/html/2609.13667#bib.bib4);[Feng et al\. 2025](https://arxiv.org/html/2609.13667#bib.bib5);[Xu et al\. 2024](https://arxiv.org/html/2609.13667#bib.bib3);[Liu et al\. 2024](https://arxiv.org/html/2609.13667#bib.bib1)\)primarily leverage general\-purpose LLMs with ReAct\-style tool\-calling frameworks\. As shown in Figure[1](https://arxiv.org/html/2609.13667#S0.F1)\(a\), these methods execute each task independently without persisting reusable experience\. However, remote sensing analysis exhibits strong demand for experience reuse\([Tang et al\. 2026](https://arxiv.org/html/2609.13667#bib.bib22)\): the same region requires cross\-seasonal monitoring, different disaster events share similar data processing pipelines, and complex domain knowledge—sensor specifications, spatial reference systems, and analytical rules—must be applied repeatedly\. These recurring analyses produce both transferable task\-planning patterns and operational experience tied to specific tools, data states, and execution constraints\. Enabling geospatial agents to accumulate, abstract, and reuse effective experience from historical executions is therefore a key challenge for building long\-term reliable geospatial intelligent systems\.

A natural resolution is to enable agents to learn from experience\([Silver and Sutton 2025](https://arxiv.org/html/2609.13667#bib.bib21)\)\. Early experience\-augmented methods focus on intra\-task reflexion \(Figure[1](https://arxiv.org/html/2609.13667#S0.F1)\(b\)\): reflecting on failures to repair the current trial, but the resulting experience is not transferred to future tasks\. This motivated inter\-task memory \(Figure[1](https://arxiv.org/html/2609.13667#S0.F1)\(c\)\): extracting cases, rules, or workflows from historical tasks to guide unseen ones\([Shinn et al\. 2023](https://arxiv.org/html/2609.13667#bib.bib9);[Gou et al\. 2024](https://arxiv.org/html/2609.13667#bib.bib10);[Zhao et al\. 2024](https://arxiv.org/html/2609.13667#bib.bib11);[Wang et al\. 2024c](https://arxiv.org/html/2609.13667#bib.bib12)\)\. However, such flat experience pools lack hierarchical representation and reliable revision mechanisms\([Qiao et al\. 2024](https://arxiv.org/html/2609.13667#bib.bib24)\)\.

A promising direction is to organize experience as*skills*—reusable procedural knowledge units with explicit applicability conditions and validation criteria\([Wang et al\. 2024a](https://arxiv.org/html/2609.13667#bib.bib14);[Mi et al\. 2026](https://arxiv.org/html/2609.13667#bib.bib15);[Jiang et al\. 2026](https://arxiv.org/html/2609.13667#bib.bib16)\)\. However, applying skill learning to geospatial agents remains challenging due to two fundamental properties of geospatial workflows\.

First, geospatial experience requires hierarchical skill representation\.Geospatial workflows contain heterogeneous knowledge with distinct transfer requirements\. High\-level planning knowledge determines*what to do*—task decomposition and tool\-chain organization—and exhibits strong cross\-task transferability \(e\.g\., flood mapping consistently follows data selection, preprocessing, water body extraction, and area estimation\)\([Qiao et al\. 2024](https://arxiv.org/html/2609.13667#bib.bib24);[Huang et al\. 2022](https://arxiv.org/html/2609.13667#bib.bib29)\)\. In contrast, tool\-level knowledge determines*how to execute*—parameter constraints, I/O dependencies, and physical feasibility conditions—and strongly depends on specific data states \(e\.g\., compatible CRS and resolution matching\)\([Wang et al\. 2024b](https://arxiv.org/html/2609.13667#bib.bib27);[Qin et al\. 2023](https://arxiv.org/html/2609.13667#bib.bib19)\)\. These two knowledge types have conflicting representation needs: excessive execution details reduce planning transferability, while excessive abstraction removes constraints for reliable tool execution\.

Second, geospatial skills require reliable evolution mechanisms\.Geospatial operations modify structured spatial states with long\-range dependencies and sometimes irreversible transformations\. Reprojection, resampling, and temporal filtering alter intermediate products and may affect all subsequent analyses\. Execution failures are difficult to attribute: an incorrect result may originate from early planning errors, invalid tool parameters, or environmental changes, with error signals manifesting only after several downstream operations\([Zhang et al\. 2025](https://arxiv.org/html/2609.13667#bib.bib25);[Zhu and others 2025](https://arxiv.org/html/2609.13667#bib.bib26)\)\. Directly asking an LLM to rewrite complete skills based on failed trajectories may overwrite previously validated components and gradually contaminate the skill bank\.

To address these challenges, we proposeGeoSkill, an experience\-driven hierarchical skill learning framework for geospatial agents\. As shown in Figure[1](https://arxiv.org/html/2609.13667#S0.F1)\(d\), GeoSkill transforms historical execution experience into reusable procedural skills through a Retrieve–ReAct–Revise–Retain loop\. GeoSkill comprises two core components: \(i\)Hierarchical Skill Bank \(HSB\), which decouples geospatial experience into Planning Skills and Tool Skills, separately modeling transferable workflow structures and executable tool constraints; and \(ii\)Causal Trace\-driven Skill Revision \(CTSR\), consisting of a Judge for failure attribution, a Critic for causal defect localization, and a Refiner for constraint\-preserving targeted revision\. Each candidate revision must pass re\-execution validation before being committed to the skill bank, preventing misattributed revisions from contaminating the skill library\. GeoSkill continuously evolves hierarchical geospatial skills without updating the underlying LLM parameters\.

Our main contributions are:

- •We proposeGeoSkill, an experience\-driven hierarchical skill learning framework that transforms historical execution experience into reusable, verifiable, and continuously evolvable procedural skills, enabling geospatial agents to accumulate capabilities across tasks without updating LLM parameters\.
- •We designHSB, which decomposes experience into planning skills and tool skills, separately modeling cross\-task transferable workflow structures and execution knowledge constrained by data states and tool specifications\.
- •We proposeCTSR, a trace\-driven skill revision mechanism that decouples failure attribution, defect localization, and targeted modification into three sequential decisions, and commits revisions only after environment re\-execution validation, preventing unreliable skill accumulation from misattribution\.

Experiments on EarthBench and ThinkGeo demonstrate that GeoSkill effectively transforms historical execution experience into reusable hierarchical skills, improving both end\-to\-end task accuracy and tool\-execution reliability\.

## Related Work

### LLM\-based Geospatial Agents

Recent advances in LLMs have driven geospatial agents from task\-specific models toward interactive systems that plan analysis procedures and invoke external tools\. Early work combined LLMs with remote sensing models: Change\-Agent integrates language models with change detection; RS\-ChatGPT and RS\-Agent enable language\-driven scene understanding and object detection\([Liu et al\. 2024](https://arxiv.org/html/2609.13667#bib.bib1);[Guo et al\. 2024](https://arxiv.org/html/2609.13667#bib.bib2);[Xu et al\. 2024](https://arxiv.org/html/2609.13667#bib.bib3)\)\. As task complexity increases, ThinkGeo and Earth\-Agent explore executable geospatial workflows for end\-to\-end analysis\([Shabbir et al\. 2025](https://arxiv.org/html/2609.13667#bib.bib4);[Feng et al\. 2025](https://arxiv.org/html/2609.13667#bib.bib5)\)\. HTAM enhances long\-horizon decomposition via hierarchical task abstraction, while CangLing\-KnowFlow and RSMeM introduce process knowledge bases and memory mechanisms for post\-failure recovery\([Li et al\. 2025](https://arxiv.org/html/2609.13667#bib.bib6);[Chen et al\. 2025](https://arxiv.org/html/2609.13667#bib.bib7);[Wu et al\. 2026](https://arxiv.org/html/2609.13667#bib.bib8)\)\. However, existing agents focus on completing individual tasks\. Since remote sensing is highly repetitive—same regions require long\-term monitoring, different events share processing pipelines, and domain knowledge must be applied repeatedly—how to abstract historical executions into reusable and improvable procedural knowledge remains unexplored\.

### Experience and Memory Learning for LLM Agents

Experience learning enables LLM agents to leverage historical interactions\. Intra\-task methods such as Reflexion and CRITIC use self\-reflection and tool feedback for error correction within the current task\([Shinn et al\. 2023](https://arxiv.org/html/2609.13667#bib.bib9);[Gou et al\. 2024](https://arxiv.org/html/2609.13667#bib.bib10)\)\. For cross\-task reuse, ExpeL extracts transferable insights from historical trajectories; AWM organizes action patterns through workflow memory; A\-MEM manages long\-term experience via structured memory\([Zhao et al\. 2024](https://arxiv.org/html/2609.13667#bib.bib11);[Wang et al\. 2024c](https://arxiv.org/html/2609.13667#bib.bib12);[Xu et al\. 2025](https://arxiv.org/html/2609.13667#bib.bib13)\)\. Tool\-augmented research further explores learning tool\-calling patterns\([Qin et al\. 2023](https://arxiv.org/html/2609.13667#bib.bib19);[Patil et al\. 2023](https://arxiv.org/html/2609.13667#bib.bib17);[Schick et al\. 2023](https://arxiv.org/html/2609.13667#bib.bib18)\)\. Recent work also separates global planning knowledge from local execution knowledge to improve agent performance\([Qiao et al\. 2024](https://arxiv.org/html/2609.13667#bib.bib24);[Wang et al\. 2024b](https://arxiv.org/html/2609.13667#bib.bib27);[Huang et al\. 2022](https://arxiv.org/html/2609.13667#bib.bib29)\)\. However, existing representations—trajectory summaries, text rules, or cases—provide semantic\-level guidance without modeling structured execution constraints\. For geospatial analysis, effective experience must encode strict constraints on data states, spatial reference systems, resolution, and tool parameters\. Moreover, existing methods lack mechanisms for diagnosing and revising erroneous or outdated knowledge\([Zhang et al\. 2025](https://arxiv.org/html/2609.13667#bib.bib25)\)\.

### Skill Learning and Agent Capability Evolution

Recent work proposes organizing experience as*skills*—procedural knowledge units with explicit triggering conditions and execution procedures\([Wang et al\. 2024a](https://arxiv.org/html/2609.13667#bib.bib14);[Mi et al\. 2026](https://arxiv.org/html/2609.13667#bib.bib15);[Jiang et al\. 2026](https://arxiv.org/html/2609.13667#bib.bib16);[Pang et al\. 2024](https://arxiv.org/html/2609.13667#bib.bib28)\)\. Beyond skill acquisition, SkillHone preserves diagnostic information and decision history during modification for progressive optimization\([Zhou and others 2026](https://arxiv.org/html/2609.13667#bib.bib23)\)\. Meanwhile, automated failure attribution has emerged as a critical challenge: even state\-of\-the\-art reasoning models achieve only 14\.2% accuracy in pinpointing failure steps\([Zhang et al\. 2025](https://arxiv.org/html/2609.13667#bib.bib25)\), and cascading failures where a single root\-cause error propagates through subsequent decisions remain a fundamental bottleneck\([Zhu and others 2025](https://arxiv.org/html/2609.13667#bib.bib26)\)\. However, existing methods primarily target general software or web tasks, treating skills as independent assets without considering domain\-specific dependencies\. For geospatial workflows, skills encompass both planning patterns and tool\-level execution constraints, and local errors may propagate through multiple intermediate products before manifesting as failures\. GeoSkill addresses these gaps by decoupling planning from tool knowledge via hierarchical representation, and performing trace\-driven defect localization, targeted modification, and re\-execution validation through CTSR\.

## Method

![Refer to caption](https://arxiv.org/html/2609.13667v1/fig2_method.png)Figure 2:Overview of the GeoSkill framework\. Duringdevelopment\(top\), GeoSkill iteratively evolves a Hierarchical Skill Bank \(HSB\) through a Retrieve–ReAct–Revise–Retain loop\. The HSB decomposes procedural memory into Planning Skills and Tool Skills\. Upon execution failure, a Causal Trace\-driven Skill Revision \(CTSR\) pipeline—Judge \(failure attribution\), Critic \(trace analysis and defect localization\), and Refiner \(constraint\-preserving targeted revision\)—diagnoses and repairs defective skills\. Revised candidates are retained only after successful re\-execution\. Duringdeployment\(bottom\), the skill bank is frozen and retrieved skills guide execution on unseen tasks without further updates\.GeoSkill transforms historical execution experience into reusable procedural skills through a closed\-loop process ofretrieval, execution, revision, and retention\. The framework operates in two stages: duringdevelopment, skills are iteratively evolved using historical tasks; duringdeployment, the learned skill bank is frozen and used for inference only\. No LLM parameters are updated in either stage\.

As shown in Fig\.[2](https://arxiv.org/html/2609.13667#Sx3.F2), GeoSkill consists of two core components: \(i\)Hierarchical Skill Bank \(HSB\), which separates planning\-level knowledge from execution\-level constraints; and \(ii\)Causal Trace\-driven Skill Revision \(CTSR\), which performs failure attribution, defect localization, and constraint\-preserving revision through a Judge–Critic–Refiner pipeline\.

### Problem Formulation

We formulate experience learning for geospatial agents as a Memory\-based Markov Decision Process \(M\-MDP\)\. Given a geospatial taskqq, the agent interacts with an executable environment through multi\-step reasoning, tool invocation, and answer generation:

𝒯M​\-​MDP=⟨𝒮,𝒜,𝒫,ℛ,γ,𝕄⟩,\\mathcal\{T\}\_\{\\mathrm\{M\\text\{\-\}MDP\}\}=\\langle\\mathcal\{S\},\\mathcal\{A\},\\mathcal\{P\},\\mathcal\{R\},\\gamma,\\mathbb\{M\}\\rangle,\(1\)where𝒮\\mathcal\{S\}and𝒜\\mathcal\{A\}denote the state and action spaces, respectively;𝒫⁡\(st\+1∣st,at\)\\mathcal\{P\}\(s\_\{t\+1\}\\mid s\_\{t\},a\_\{t\}\)is the environment transition kernel;ℛ\\mathcal\{R\}denotes execution feedback;γ\\gammais the discount factor; and𝕄\\mathbb\{M\}is the space of external memories\. A geospatial statests\_\{t\}contains the task objective, available data, previous tool calls, intermediate artifacts, and environment feedback\. An actionata\_\{t\}may correspond to a reasoning decision, a parameterized tool invocation, or a final answer\.

GeoSkill instantiates external procedural memory as a skill bankMr∈𝕄M^\{r\}\\in\\mathbb\{M\}, whererrindexes the rollout for the current task\. Rather than repeatedly changing the retrieved context at every decision step, GeoSkill selects a stable skill context at the beginning of each rollout:

Sr∼μ⁡\(S∣q,Mr\),S^\{r\}\\sim\\mu\(S\\mid q,M^\{r\}\),\(2\)whereμ\\muis a hierarchical retrieval function andSrS^\{r\}contains one planning skill together with the tool skills required by its workflow\. The retrieved skills condition action generation throughout the ReAct trajectory:

atr∼pθ​\(at∣str,q,Sr\),a\_\{t\}^\{r\}\\sim p\_\{\\theta\}\(a\_\{t\}\\mid s\_\{t\}^\{r\},q,S^\{r\}\),\(3\)wherepθp\_\{\\theta\}is the frozen underlying LLM\. A rollout produces

τr=\{s0r,a0r,o0r,…,sTr,aTr,oTr\},\\tau^\{r\}=\\\{s\_\{0\}^\{r\},a\_\{0\}^\{r\},o\_\{0\}^\{r\},\\ldots,s\_\{T\}^\{r\},a\_\{T\}^\{r\},o\_\{T\}^\{r\}\\\},\(4\)and a trajectory\-level verifier𝒢⁡\(τr,y∗\)∈\{0,1\}\\mathcal\{G\}\(\\tau^\{r\},y^\{\*\}\)\\in\\\{0,1\\\}determines task success, wherey∗y^\{\*\}is a semantic reference answer\. When available, a reference explanation or trajectoryτ∗\\tau^\{\*\}is used as diagnostic evidence but is not treated as the only valid execution path\.

GeoSkill does not optimizeθ\\theta\. Instead, it updates external procedural memory at the rollout level:

Mr\+1=𝒰⁡\(Mr,Xr\),Xr=\(q,Sr,τr,ϵr,y^r,y∗,τ∗\),M^\{r\+1\}=\\mathcal\{U\}\(M^\{r\},X^\{r\}\),\\quad X^\{r\}=\(q,S^\{r\},\\tau^\{r\},\\epsilon^\{r\},\\hat\{y\}^\{r\},y^\{\*\},\\tau^\{\*\}\),\(5\)whereϵr\\epsilon^\{r\}denotes environment outputs or errors,y^r\\hat\{y\}^\{r\}is the predicted answer, and𝒰\\mathcal\{U\}is instantiated by the proposed CTSR mechanism\.

### Overview of GeoSkill

As illustrated in Fig\.[2](https://arxiv.org/html/2609.13667#Sx3.F2), GeoSkill follows a Retrieve–ReAct–Revise–Retain lifecycle during development\. It first analyzes the task intent, retrieves a planning skill from HSB, and then retrieves the tool skills required by the tool chain encoded in the planning skill\. When a required skill is absent, the LLM initializes a candidate skill\. The resulting hierarchical skill bundle is injected into a ReAct agent\. A successful rollout validates the current candidate bundle and allows it to be retained\. For an unsuccessful rollout, CTSR reads the complete evidence and distinguishes skill defects from agent execution deviations and runtime failures\. Only failures attributed to skill knowledge trigger skill revision\. After all defective skills have been processed, the complete candidate bundle is re\-executed\. The candidates are committed to HSB only if the new rollout succeeds\. Otherwise, revision continues within a bounded rollout budget, after which all task\-specific candidates are discarded and the pre\-task skill bank is restored\.

### Hierarchical Skill Bank

Geospatial executions contain two types of procedural knowledge with different abstraction requirements\. Workflow\-level knowledge should abstract reusable task structures from specific regions, sensors, and parameter settings\. Tool\-level knowledge must instead preserve constraints imposed by input data, intermediate artifacts, and the execution environment\. Mixing them in a single skill either binds planning knowledge to instance\-specific configurations or removes the physical and geometric conditions required for reliable execution\. GeoSkill therefore decomposes procedural memory as

Mr=\(𝒦r,ℰr\),M^\{r\}=\(\\mathcal\{K\}^\{r\},\\mathcal\{E\}^\{r\}\),\(6\)where𝒦r\\mathcal\{K\}^\{r\}andℰr\\mathcal\{E\}^\{r\}denote the Planning Skill Bank and Tool Skill Bank, respectively\. Both skill types follow a shared structured record while differing in granularity and retrieval:

s=\(id,d,ω,W,V,T\),s=\(\\mathrm\{id\},d,\\omega,W,V,T\),\(7\)whereid\\mathrm\{id\}is the skill identifier,ddis its functional description,ω\\omegaspecifies when it should be used,WWis an executable workflow or invocation template,VVcontains validation constraints, andTTdefines termination conditions\. In a planning skill,WWdescribes a multi\-step tool chain and its dependencies\. In a tool skill,WWdescribes the parameter schema, input–output contract, and execution notes of one tool\.

#### Planning Skill Retrieval and Initialization

Planning skills represent task\-level procedural knowledge, including task decomposition, subgoal ordering, and tool\-chain topology\. Given a taskqq, GeoSkill retrieves the most applicable planning skill from𝒦r\\mathcal\{K\}^\{r\}:

kr=μp​\(q,𝒦r\)\.k^\{r\}=\\mu\_\{p\}\(q,\\mathcal\{K\}^\{r\}\)\.\(8\)A planning skill abstracts the stable workflow structure of a task family\. Its workflow template specifies required tools, stage dependencies, and validation requirements, and serves as a Plan\-and\-Execute prior for the subsequent ReAct rollout\. If retrieval returns no applicable skill, the system initializes a candidate planning skillk~r\\tilde\{k\}^\{r\}from the task description and available tool space\. The candidate remains in a task\-local working copy and must pass rollout verification before it can be retained\.

#### Tool Skill Retrieval and Initialization

GeoSkill parses the workflow template ofkrk^\{r\}to obtain its tool chain

𝒯⁡\(kr\)=\(u1,…,uL\),\\mathcal\{T\}\(k^\{r\}\)=\(u\_\{1\},\\ldots,u\_\{L\}\),\(9\)whereuju\_\{j\}denotes thejj\-th tool operation\. The corresponding tool skill is retrieved fromℰr\\mathcal\{E\}^\{r\}using the tool identifier:

ejr=μt​\(uj,ℰr\)\.e\_\{j\}^\{r\}=\\mu\_\{t\}\(u\_\{j\},\\mathcal\{E\}^\{r\}\)\.\(10\)Tool skills preserve executable knowledge for individual operators, including parameter names and types, input requirements, output semantics, artifact dependencies, spatial\-reference and resolution constraints, and validation conditions\. If a required tool skill is missing, GeoSkill initializes a candidatee~jr\\tilde\{e\}\_\{j\}^\{r\}\. The activated hierarchical skill bundle is

Sr=\(kr,ℰ⁡\(kr\)\),ℰ⁡\(kr\)=\{ejr\}j=1L\.S^\{r\}=\(k^\{r\},\\mathcal\{E\}\(k^\{r\}\)\),\\quad\\mathcal\{E\}\(k^\{r\}\)=\\\{e\_\{j\}^\{r\}\\\}\_\{j=1\}^\{L\}\.\(11\)This tool\-chain\-conditioned retrieval avoids indiscriminate retrieval from the complete tool bank and keeps the injected tool knowledge consistent with the selected workflow\.

### Skill\-Conditioned ReAct

GeoSkill injects the complete skill bundleSrS^\{r\}as a stable procedural context for one ReAct rollout\. The planning skill provides a workflow\-level prior, while the tool skills constrain parameter instantiation and artifact propagation\. The agent remains responsive to environment observations, but its tool choices and parameter generation are explicitly guided by the activated skills\. After execution, GeoSkill records the task, rendered skill contents, ReAct trajectory, skill\-compliance report, environment outputs or errors, predicted answer, and semantic reference answer, forming the diagnostic evidenceXrX^\{r\}in Eq\. \([5](https://arxiv.org/html/2609.13667#Sx3.E5)\)\. If𝒢⁡\(τr,y∗\)=1\\mathcal\{G\}\(\\tau^\{r\},y^\{\*\}\)=1, the bundle proceeds to Retain\. Otherwise, the evidence is passed to CTSR\.

### Causal Trace\-Driven Skill Revision

A failed geospatial rollout does not directly imply that the underlying skills are defective\. The same final error may arise from an incomplete planning skill, an invalid tool parameter, failure to follow correct skills, incorrect answer synthesis, or a runtime exception\. Blindly rewriting skills from every failed trajectory can attribute execution errors to valid knowledge and overwrite previously verified components\. CTSR separates responsibility routing, skill\-level diagnosis, and constrained revision\. These stages are implemented through sequential role prompts within one shared LLM conversation context, avoiding information loss from summarizing evidence across independent agents\.

#### Judge: Failure Responsibility Routing

Judge receives the complete evidenceXrX^\{r\}and outputs

Jr=\(ρr,Ir,Pr,Er,zr\),J^\{r\}=\(\\rho^\{r\},I^\{r\},P^\{r\},E^\{r\},z^\{r\}\),\(12\)whereρr\\rho^\{r\}is the responsibility category,IrI^\{r\}is the set of skills to inspect,PrP^\{r\}is the preserve ledger,ErE^\{r\}contains factual evidence, andzrz^\{r\}is a concise responsibility rationale\. The responsibility space distinguishes planning\-skill defects, tool\-skill defects, joint skill defects, agent non\-compliance, answer synthesis errors, runtime errors, and successful execution\. Judge only decides whether skills should be edited and which skills require inspection\. It does not produce detailed diagnoses or patches\. For agent non\-compliance or answer synthesis errors, GeoSkill generates an ephemeral verbal reflection for the next rollout without modifying HSB\. The preserve ledgerPrP^\{r\}identifies skill components already supported by the execution evidence; each entry specifies a target skill, an internal reference, and the reason why that component must remain unchanged\.

#### Critic: Skill\-Level Causal Diagnosis

For each target skillsi∈Irs\_\{i\}\\in I^\{r\}, Critic performs an independent, read\-only diagnosis:

Dir=\(Fir,Eir,Cir,Lir,Zir,Pir\),D\_\{i\}^\{r\}=\(F\_\{i\}^\{r\},E\_\{i\}^\{r\},C\_\{i\}^\{r\},L\_\{i\}^\{r\},Z\_\{i\}^\{r\},P\_\{i\}^\{r\}\),\(13\)whereFirF\_\{i\}^\{r\}denotes the defect family,EirE\_\{i\}^\{r\}contains skill\-specific evidence,CirC\_\{i\}^\{r\}expresses the causal chain from the earliest skill defect to the downstream failure,LirL\_\{i\}^\{r\}specifies the fields or fragments that require rewriting,ZirZ\_\{i\}^\{r\}is a diagnostic summary, andPirP\_\{i\}^\{r\}extends the preserve ledger when additional validated components are identified\. For planning skills, Critic examines missing stages, incorrect tool selection, invalid operation ordering, broken artifact dependencies, and termination logic\. For tool skills, it examines parameter names and values, tool contracts, input–output objects, data dependencies, physical constraints, and metric\-to\-answer mapping\.

#### Refiner: Constraint\-Preserving Targeted Revision

Refiner reads the shared execution evidence, the Judge output, and the Critic diagnosis, and generates a structured patch:

\(Zi,revr,Δir\)=ℛ⁡\(Xr,Jr,Dir\),\(Z\_\{i,\\mathrm\{rev\}\}^\{r\},\\Delta\_\{i\}^\{r\}\)=\\mathcal\{R\}\(X^\{r\},J^\{r\},D\_\{i\}^\{r\}\),\(14\)whereZi,revrZ\_\{i,\\mathrm\{rev\}\}^\{r\}records the verifier contract, revision evidence, and expected execution anchors, whileΔir\\Delta\_\{i\}^\{r\}is the actual patch set\. Each patch identifies the target skill, target field, operation, and replacement payload, rather than regenerating the entire skill\. Preservation is enforced through a deterministic patch guard\. LetRef⁡\(δ\)\\operatorname\{Ref\}\(\\delta\)denote the internal skill references touched by a patchδ\\delta\. A replacement or deletion patch is rejected whenever it intersects the preserve ledger without explicit evidence of a direct conflict:

Accept\(δ\)=0,ifop\(δ\)∈\{replace,delete\},Ref\(δ\)∩Pir≠∅,conflict\(δ\)=0\.\\operatorname\{Accept\}\(\\delta\)=0,\\;\\text\{if \}\\operatorname\{op\}\(\\delta\)\\in\\\{\\mathrm\{replace\},\\mathrm\{delete\}\\\},\\\\ \\operatorname\{Ref\}\(\\delta\)\\cap P\_\{i\}^\{r\}\\neq\\varnothing,\\;\\operatorname\{conflict\}\(\\delta\)=0\.\(15\)When Judge identifies multiple defective skills, they are diagnosed and revised sequentially in the same shared context\. The resulting candidate bundle is

S~r\+1=Apply⁡\(Sr,\{Δir\}si∈Ir\)\.\\widetilde\{S\}^\{r\+1\}=\\operatorname\{Apply\}\(S^\{r\},\\\{\\Delta\_\{i\}^\{r\}\\\}\_\{s\_\{i\}\\in I^\{r\}\}\)\.\(16\)

### Retain and Development Procedure

Newly initialized and revised skills are maintained in a task\-local working copy\. A candidate is never committed merely because the LLM judges it to be plausible\. Instead, the complete candidate bundle is used to re\-execute the original task:

τr\+1=ReAct⁡\(q,S~r\+1\)\.\\tau^\{r\+1\}=\\operatorname\{ReAct\}\(q,\\widetilde\{S\}^\{r\+1\}\)\.\(17\)If𝒢⁡\(τr\+1,y∗\)=1\\mathcal\{G\}\(\\tau^\{r\+1\},y^\{\*\}\)=1, all newly initialized or revised skills in the working copy are atomically committed to HSB\. Otherwise, CTSR processes the new trajectory in the next rollout\. Development uses a bounded rollout budgetBB\. If no rollout succeeds, all candidates and patches produced for the task are discarded and the pre\-task bank snapshot is preserved\. Retain therefore serves jointly as an execution\-verification gate and a transactional commit mechanism, preventing partial updates from breaking consistency between planning and tool skills\.

### Deployment with a Frozen Skill Bank

After development, GeoSkill freezes the hierarchical skill bank:

M∗=\(𝒦∗,ℰ∗\)\.M^\{\*\}=\(\\mathcal\{K\}^\{\*\},\\mathcal\{E\}^\{\*\}\)\.\(18\)For an unseen task, the system only performs intent analysis, planning skill retrieval, tool\-chain parsing, tool skill retrieval, and skill\-conditioned ReAct\. Deployment has no access to reference answers, does not invoke Judge, Critic, or Refiner, and neither generates nor writes back skills\. If no applicable planning skill is retrieved, the system falls back to vanilla ReAct without skill assistance\.

## Experiments

### Experimental Setup

Benchmarks\.We evaluate GeoSkill on two representative geospatial agent benchmarks: EarthBench and ThinkGeo\.

EarthBench focuses on end\-to\-end geospatial workflow execution, including tool selection, parameter configuration, and final answer generation\. ThinkGeo further decomposes geospatial agent ability into perception, orchestration, and logical reasoning\.

To evaluate cross\-task experience transfer, both benchmarks are divided into two balanced partitions following their original task distributions\. Experience\-based methods, including ExpeL, AWM, and GeoSkill, first construct experience or skills from the first partition and then perform frozen evaluation on the second partition\. Reasoning\-based baselines, including ReAct and Reflexion, directly solve the evaluation partition without prior experience accumulation\.

Baselines\.We compare GeoSkill with representative reasoning and experience\-learning agents\.

ReAct is adopted as the standard tool\-augmented reasoning baseline\. ReAct×3\\times 3repeatedly executes the same task to examine whether additional rollouts can substitute experience reuse\. Reflexion improves execution through verbal self\-correction\.

ExpeL and AWM represent inter\-task experience memory approaches\. To investigate whether better experience extraction alone can improve transfer, we evaluate†\\daggervariants \(ExpeL†, AWM†, GeoSkill†\), where experiences or skills are constructed using DeepSeek\-V4\-Pro while execution is performed by the target backbone\.

### Evaluation Metrics

We evaluate geospatial agents from three dimensions:Task success is measured by end\-to\-end Accuracy \(Acc\)\. Following EarthBench, tool execution reliability is evaluated using Tool\-Any, Tool\-In, Tool\-Exact, and Parameter Accuracy, which measure tool selection, workflow alignment, exact tool matching, and parameter correctness\. Efficiency \(Eff\.\) measures execution redundancy through the ratio between predicted and reference tool calls\. For ThinkGeo, we report Perception \(P\), Orchestration \(O\), Logic \(L\), and answer accuracy following the original benchmark protocol\.

### Main Results and Analysis

Table 1:Main results on EarthBench and ThinkGeo\.Red= best per backbone,blue= second best\.Table[1](https://arxiv.org/html/2609.13667#Sx4.T1)summarizes the overall comparison on EarthBench and ThinkGeo\.

GeoSkill improves both task success and execution reliability\.On EarthBench with DeepSeek\-V4\-Pro, GeoSkill achieves 79\.03% accuracy, outperforming ReAct \(58\.06%\) by 20\.97 points and Reflexion \(73\.98%\) by 5\.05 points\. Meanwhile, GeoSkill reduces the efficiency ratio from 12\.93 \(Reflexion\) to 6\.08, close to the ground\-truth baseline, demonstrating that improvement comes from better planning rather than trial\-and\-error\. In contrast, ReAct@3 improves accuracy to 66\.94% but increases redundancy to 12\.85, confirming that repeated attempts without structured memory are inefficient\.

Hierarchical skills compensate for weaker backbones\.The advantage of GeoSkill becomes more pronounced on smaller models\. With Qwen3\-8B, GeoSkill reaches 41\.13% accuracy, improving over ReAct \(11\.29%\) by 29\.84 points and over Reflexion \(37\.10%\) by 4\.03 points\. More importantly, Tool\-Exact increases from 8\.47% \(Reflexion\) to 22\.59%, and Parameter Accuracy from 5\.09% to 13\.88%\. These gains are due to HSB’s explicit modeling of tool\-level constraints—parameter names, types, and dependencies—which are precisely where small models struggle\.

Better experience extraction alone is insufficient\.To isolate the effect of extraction quality from representation design, we evaluate†\\daggervariants where experiences or skills are constructed using DeepSeek\-V4\-Pro while execution is performed by the target backbone\. On Qwen3\-32B, ExpeL†improves from 17\.74% to 21\.77%, and GeoSkill†further improves GeoSkill from 41\.13% to 43\.55%, confirming that stronger extraction benefits both flat and hierarchical representations\. However, on Qwen3\-8B, GeoSkill†\(38\.71%\) underperforms GeoSkill \(41\.13%\)\. This reversal occurs because skills constructed by a much stronger model encode complex multi\-step reasoning patterns that exceed the 8B model’s execution capacity, causing more frequent deviations from the intended workflow\. In contrast, skills evolved by the same backbone better match its reasoning budget\.

Transferable skills improve geospatial reasoning\.On ThinkGeo, GeoSkill achieves consistent improvements in orchestration and logic capabilities\. With DeepSeek\-V4\-Pro, Logic F1 improves from 41\.88% \(ReAct\) to 59\.61%, a 17\.73\-point gain\. On Qwen3\-8B, GeoSkill improves Perception from 48\.73% to 53\.40% \(\+4\.67\), Operation from 43\.14% to 47\.82% \(\+4\.68\), and Logic from 29\.49% to 33\.60% \(\+4\.11\)\. These improvements demonstrate that hierarchical skills capture reusable multi\-step reasoning patterns that transfer across tasks\.

### Ablation Study

Table 2:Ablation of GeoSkill components \(EarthBench, DSV4\-Pro\)\.All ablations are conducted on EarthBench with DeepSeek\-V4\-Pro as the backbone\. We analyze the contribution of each core component\. The ablation results demonstrate that hierarchical skill representation provides the foundation for experience transfer\. Introducing HSB improves accuracy from 65\.32% to 72\.58%, indicating that separating planning knowledge from tool constraints enables more effective reuse\. Adding CTSR further improves performance to 79\.03%\. This gain verifies that skill evolution requires reliable failure diagnosis and selective revision rather than directly rewriting stored experiences\.

B=0B\{=\}0B=1B\{=\}1B=2B\{=\}250505555606065657070757580808585Revision Round \(BB\)Score \(%\)AccuracyTool\-ExactFigure 3:Effect of revision rounds on EarthBench \(DeepSeek\-V4\-Pro\)\.B=0B\{=\}0is retrieval\-only without revision\.
### Skill Evolution Analysis

We further analyze how iterative skill refinement affects performance\.As shown in Figure[3](https://arxiv.org/html/2609.13667#Sx4.F3), performance improves monotonically fromB=0B\{=\}0\(72\.58%\) toB=2B\{=\}2\(79\.03%\), with the first revision round contributing the largest accuracy gain \(\+3\.23%\)\. Tool\-Exact also improves steadily from 55\.80% to 58\.37%, confirming that CTSR refines both planning and tool\-level knowledge\. TheB=0B\{=\}0result already outperforms ReAct \(58\.06%\), confirming that the skill bank provides value even before evolution\.

## Conclusion

We presented GeoSkill, an experience\-driven hierarchical skill learning framework that enables geospatial agents to distill prior executions into reusable procedural knowledge\. HSB decouples planning from tool constraints for transferable yet executable skills, while CTSR performs attribution, localization, and validated patching to prevent unreliable accumulation\. Experiments on EarthBench and ThinkGeo confirm consistent improvements in accuracy and reliability\.

## References

- Chenet al\.\(2025\)Z\. Chen, H\. Wang, J\. Yao, J\. Zhang, P\. Ghamisi, J\. Zhou, P\. M\. Atkinson, and B\. ZhangCangLing\-KnowFlow: a unified knowledge\-and\-flow\-fused agent for comprehensive remote sensing applications\.arXiv preprint arXiv:2512\.15231\.Cited by:[LLM\-based Geospatial Agents](https://arxiv.org/html/2609.13667#Sx2.SSx1.p1.1)\.
- Fenget al\.\(2025\)P\. Feng, Z\. Lv, J\. Ye, X\. Wang, X\. Huo, J\. Yu, W\. Xu, W\. Zhang, L\. Bai, C\. He, and W\. LiEarth\-Agent: unlocking the full landscape of earth observation with agents\.arXiv preprint arXiv:2509\.23141\.Note:Published at ICLR 2026Cited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p1.1),[LLM\-based Geospatial Agents](https://arxiv.org/html/2609.13667#Sx2.SSx1.p1.1)\.
- Gouet al\.\(2024\)Z\. Gou, Z\. Shao, Y\. Gong, Y\. Shen, Y\. Yang, N\. Duan, and W\. ChenCRITIC: large language models can self\-correct with tool\-interactive critiquing\.InThe Twelfth International Conference on Learning Representations,Cited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p2.1),[Experience and Memory Learning for LLM Agents](https://arxiv.org/html/2609.13667#Sx2.SSx2.p1.1)\.
- Guoet al\.\(2024\)H\. Guo, X\. Su, C\. Wu, B\. Du, L\. Zhang, and D\. LiRemote sensing chatgpt: solving remote sensing tasks with chatgpt and visual models\.arXiv preprint arXiv:2401\.09083\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2401.09083)Cited by:[LLM\-based Geospatial Agents](https://arxiv.org/html/2609.13667#Sx2.SSx1.p1.1)\.
- Huanget al\.\(2022\)W\. Huang, F\. Xia, T\. Xiao, H\. Chan, J\. Liang, P\. Florence, A\. Zeng, J\. Tompson, I\. Mordatch, Y\. Chebotar, P\. Sermanet, T\. Jackson, N\. Brown, L\. Luu, S\. Levine, K\. Hausman, and B\. IchterInner monologue: embodied reasoning through planning with language models\.arXiv preprint arXiv:2207\.05608\.Cited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p4.1),[Experience and Memory Learning for LLM Agents](https://arxiv.org/html/2609.13667#Sx2.SSx2.p1.1)\.
- Jianget al\.\(2026\)G\. Jiang, Z\. Su, X\. Qu, and Y\. R\. FungXSkill: continual learning from experience and skills in multimodal agents\.arXiv preprint arXiv:2603\.12056\.Cited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p3.1),[Skill Learning and Agent Capability Evolution](https://arxiv.org/html/2609.13667#Sx2.SSx3.p1.1)\.
- Liet al\.\(2025\)K\. Li, J\. Wang, Z\. Wang, H\. Qiao, W\. Zhang, D\. Meng, and X\. CaoDesigning domain\-specific agents via hierarchical task abstraction mechanism\.arXiv preprint arXiv:2511\.17198\.Cited by:[LLM\-based Geospatial Agents](https://arxiv.org/html/2609.13667#Sx2.SSx1.p1.1)\.
- Liuet al\.\(2024\)C\. Liu, K\. Chen, H\. Zhang, Z\. Qi, Z\. Zou, and Z\. ShiChange\-agent: towards interactive comprehensive remote sensing change interpretation and analysis\.IEEE Transactions on Geoscience and Remote Sensing62,pp\. 1–17\.External Links:[Document](https://dx.doi.org/10.1109/TGRS.2024.3425815)Cited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p1.1),[LLM\-based Geospatial Agents](https://arxiv.org/html/2609.13667#Sx2.SSx1.p1.1)\.
- Miet al\.\(2026\)Q\. Mi, Z\. Ma, M\. Yang, H\. Li, Y\. Wang, H\. Zhang, and J\. WangSkill\-Pro: learning reusable skills from experience via non\-parametric PPO for LLM agents\.InProceedings of the 43rd International Conference on Machine Learning,Note:Spotlight presentationCited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p3.1),[Skill Learning and Agent Capability Evolution](https://arxiv.org/html/2609.13667#Sx2.SSx3.p1.1)\.
- Muniret al\.\(2026\)M\. A\. Munir, M\. U\. Sheikh, A\. Shabbir, M\. H\. Khan, F\. Khan, X\. X\. Zhu, B\. Demir, and S\. KhanAgentic AI for remote sensing: technical challenges and research directions\.arXiv preprint arXiv:2604\.24919\.Cited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p1.1)\.
- Panget al\.\(2024\)J\. Pang, S\. Yang, K\. Li, J\. Zhang, X\. Chen, N\. Tang, and Y\. YuKALM: knowledgeable agents by offline reinforcement learning from large language model rollouts\.InAdvances in Neural Information Processing Systems,Vol\.37\.Cited by:[Skill Learning and Agent Capability Evolution](https://arxiv.org/html/2609.13667#Sx2.SSx3.p1.1)\.
- Patilet al\.\(2023\)S\. G\. Patil, T\. Zhang, X\. Wang, and J\. E\. GonzalezGorilla: large language model connected with massive apis\.arXiv preprint arXiv:2305\.15334\.Cited by:[Experience and Memory Learning for LLM Agents](https://arxiv.org/html/2609.13667#Sx2.SSx2.p1.1)\.
- Qiaoet al\.\(2024\)S\. Qiao, R\. Fang, N\. Zhang, Y\. Zhu, X\. Chen, S\. Deng, Y\. Jiang, P\. Xie, F\. Huang, and H\. ChenAgent planning with world knowledge model\.InAdvances in Neural Information Processing Systems,Vol\.37,pp\. 114843–114871\.Cited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p2.1),[Introduction](https://arxiv.org/html/2609.13667#Sx1.p4.1),[Experience and Memory Learning for LLM Agents](https://arxiv.org/html/2609.13667#Sx2.SSx2.p1.1)\.
- Qinet al\.\(2023\)Y\. Qin, S\. Liang, Y\. Ye, L\. Zhu, Y\. Wu, X\. Yan, Z\. Wang, X\. Cong, Y\. Tang, Q\. Qian,et al\.ToolLLM: facilitating large language models to master 16000\+ real\-world apis\.arXiv preprint arXiv:2307\.16789\.Cited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p4.1),[Experience and Memory Learning for LLM Agents](https://arxiv.org/html/2609.13667#Sx2.SSx2.p1.1)\.
- Schicket al\.\(2023\)T\. Schick, J\. Dwivedi\-Yu, R\. Dessì, R\. Raileanu, M\. Lomeli, E\. Hambro, L\. Zettlemoyer, N\. Cancedda, and T\. ScialomToolformer: language models can teach themselves to use tools\.Advances in Neural Information Processing Systems36\.Cited by:[Experience and Memory Learning for LLM Agents](https://arxiv.org/html/2609.13667#Sx2.SSx2.p1.1)\.
- Shabbiret al\.\(2025\)A\. Shabbir, M\. A\. Munir, A\. Dudhane, M\. U\. Sheikh, M\. H\. Khan, P\. Fraccaro, J\. Bernabe Moreno, F\. S\. Khan, and S\. KhanThinkGeo: evaluating tool\-augmented agents for remote sensing tasks\.arXiv preprint arXiv:2505\.23752\.Cited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p1.1),[LLM\-based Geospatial Agents](https://arxiv.org/html/2609.13667#Sx2.SSx1.p1.1)\.
- Shinnet al\.\(2023\)N\. Shinn, F\. Cassano, E\. Berman, A\. Gopinath, K\. Narasimhan, and S\. YaoReflexion: language agents with verbal reinforcement learning\.InAdvances in Neural Information Processing Systems,Vol\.36\.Cited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p2.1),[Experience and Memory Learning for LLM Agents](https://arxiv.org/html/2609.13667#Sx2.SSx2.p1.1)\.
- Silver and Sutton \(2025\)D\. Silver and R\. S\. SuttonWelcome to the era of experience\.In Designing an Intelligence, MIT Press\.Note:PreprintCited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p2.1)\.
- Tanget al\.\(2026\)J\. Tang, Y\. Yan, Q\. Wang, Y\. Xia, B\. Geng, J\. Chen, K\. Ma, Y\. Zhai, Q\. He, W\. Shao, Y\. Sun, J\. Dai, C\. Chen, X\. Xu, K\. Yao, L\. Zhang, W\. Wei, Q\. Chen, A\. Plaza, and Y\. ZhangIntelligent remote sensing agents: a survey\.Note:PreprintCited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p1.1)\.
- Wanget al\.\(2024a\)G\. Wang, Y\. Xie, Y\. Jiang, A\. Mandlekar, C\. Xiao, Y\. Zhu, L\. Fan, and A\. AnandkumarVoyager: an open\-ended embodied agent with large language models\.InAdvances in Neural Information Processing Systems,Vol\.37\.Cited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p3.1),[Skill Learning and Agent Capability Evolution](https://arxiv.org/html/2609.13667#Sx2.SSx3.p1.1)\.
- Wanget al\.\(2024b\)X\. Wang, Y\. Chen, L\. Yuan, Y\. Zhang, Y\. Li, H\. Peng, and H\. JiExecutable code actions elicit better LLM agents\.InProceedings of the 41st International Conference on Machine Learning,Cited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p4.1),[Experience and Memory Learning for LLM Agents](https://arxiv.org/html/2609.13667#Sx2.SSx2.p1.1)\.
- Wanget al\.\(2024c\)Z\. Z\. Wang, J\. Mao, D\. Fried, and G\. NeubigAgent workflow memory\.arXiv preprint arXiv:2409\.07429\.Cited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p2.1),[Experience and Memory Learning for LLM Agents](https://arxiv.org/html/2609.13667#Sx2.SSx2.p1.1)\.
- Wuet al\.\(2026\)B\. Wu, Y\. Zhang, Z\. Guo, T\. Liu, C\. Qian, Y\. Lu, X\. Du, Y\. Li, Y\. Zhang, C\. Chen, L\. Yao, C\. Zhou, and M\. SunRSMeM: knowledge\-enhanced memory evolution for remote sensing agents with systematic evaluation\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),External Links:[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.1519)Cited by:[LLM\-based Geospatial Agents](https://arxiv.org/html/2609.13667#Sx2.SSx1.p1.1)\.
- Xuet al\.\(2024\)W\. Xu, Z\. Yu, B\. Mu, J\. Wang, Z\. Wei, and M\. PengRS\-Agent: automating remote sensing tasks through intelligent agent\.arXiv preprint arXiv:2406\.07089\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2406.07089)Cited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p1.1),[LLM\-based Geospatial Agents](https://arxiv.org/html/2609.13667#Sx2.SSx1.p1.1)\.
- Xuet al\.\(2025\)W\. Xu, Z\. Liang, K\. Mei, H\. Gao, J\. Tan, and Y\. ZhangA\-MEM: agentic memory for LLM agents\.Advances in Neural Information Processing Systems\.Cited by:[Experience and Memory Learning for LLM Agents](https://arxiv.org/html/2609.13667#Sx2.SSx2.p1.1)\.
- Zhanget al\.\(2025\)S\. Zhang, M\. Yin, J\. Zhang, J\. Liu, Z\. Han, J\. Zhang, B\. Li, C\. Wang, H\. Wang, and Y\. ChenWhich agent causes task failures and when? on automated failure attribution of LLM multi\-agent systems\.InProceedings of the 42nd International Conference on Machine Learning,Note:SpotlightCited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p5.1),[Experience and Memory Learning for LLM Agents](https://arxiv.org/html/2609.13667#Sx2.SSx2.p1.1),[Skill Learning and Agent Capability Evolution](https://arxiv.org/html/2609.13667#Sx2.SSx3.p1.1)\.
- Zhaoet al\.\(2024\)A\. Zhao, D\. Huang, Q\. Xu, M\. Lin, Y\. Liu, and G\. HuangExpeL: llm agents are experiential learners\.InProceedings of the 38th Annual AAAI Conference on Artificial Intelligence,Vol\.38,pp\. 19632–19642\.Cited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p2.1),[Experience and Memory Learning for LLM Agents](https://arxiv.org/html/2609.13667#Sx2.SSx2.p1.1)\.
- Zhouet al\.\(2026\)J\. Zhouet al\.SkillHone: a harness for continual agent skill evolution through persistent decision history\.arXiv preprint arXiv:2606\.08671\.Cited by:[Skill Learning and Agent Capability Evolution](https://arxiv.org/html/2609.13667#Sx2.SSx3.p1.1)\.
- Zhuet al\.\(2025\)Z\. Zhuet al\.Where LLM agents fail and how they can learn from failures\.arXiv preprint arXiv:2509\.25370\.Cited by:[Introduction](https://arxiv.org/html/2609.13667#Sx1.p5.1),[Skill Learning and Agent Capability Evolution](https://arxiv.org/html/2609.13667#Sx2.SSx3.p1.1)\.