PaperlayerView exact revision on arXiv

Source-linked AI summary

Earth-Agent-Pro: Towards Real-World Full-Chain Earth Observation with Agents

Zhutao Lv, Chenhao Dang, Yi Feng, Yanpei Gong, Xiaolei Wang, Junyan Ye, Conghui He, Weijia Li

arXiv:2609.12533v1cs.CVcs.AIcs.CL

TL;DR

Earth observation agents must turn high-level questions into executable workflows, but existing agents and benchmarks leave full-chain runtime execution largely untested. Earth-Agent-Pro combines skill-guided planning, evidence-linked workflow memory, and localized repair, evaluated with Earth-Bench-Pro’s matched regimes. On Earth-Bench-OW, it achieves 66.13% LLM-as-Judge accuracy with GPT-5, exceeding ReAct by 20.95 points.

  • Problem

    Existing EO agents typically start from supplied observations and benchmarks often provide prepared inputs or candidate answers, leaving full-chain open-world EO execution largely untested.

    Source: §I · pp. 1–2

  • Method

    Earth-Agent-Pro uses expert-authored skills and workflow-centered memory to guide planning, runtime operation grounding, and repair of only the affected workflow suffix.

    Source: §I · p. 2, A. Framework Overview · p. 5, D. Evidence-Grounded Execution and… · p. 7

  • Results

    66.13% LLM-as-Judge accuracy on Earth-Bench-OW with GPT-5 exceeds ReAct by 20.95 points; Earth-Agent-Pro also leads ReAct by 24.44 points on Tools-In-Order.

    Source: §I · p. 2

  • Takeaways & Limitations

    Earth-Bench-Pro evaluates matched instruction-following, autonomous-planning, and open-world execution regimes through executable evidence chains.

    Source: §I · p. 2, §VIII · p. 15

    On Product tasks, Codex reaches 75.00% accuracy versus 55.00% for the strongest Earth-Agent-Pro variant, limiting claims of uniform superiority across task families.

    Source: B. Limitations and Trade-offs · p. 15

Abstract

from arXiv · show

Real-world Earth observation (EO) agents must translate high-level scientific questions into executable workflows to acquire observations, prepare data, perform domain computations, and derive conclusions from runtime evidence. Existing EO agents typically start from supplied observations, while benchmarks typically provide prepared inputs or candidate answers, leaving full-chain open-world EO execution largely untested. We present Earth-Agent-Pro, an execution-adaptive Plan-and-Execute framework using expert-authored skills to constrain planning and runtime tool use. Workflow-centered structured memory records planned steps, accepted evidence, and their dependencies, enabling repair of only the affected workflow suffix when runtime evidence invalidates a step. Separate large language model adapters use sequence-level supervised fine-tuning for planner workflow composition and node-level group relative policy optimization with locally verifiable rewards for executor tool-argument grounding. Earth-Bench-Pro instantiates 248 expert-curated task cores as 744 questions under three matched regimes. Its 248 Open-World Execution questions span RGB imagery, spectral observations, and remote sensing products, pairing high-level requests with runtime data requirements, executable trajectories, and open-ended answers grounded in execution evidence. With a shared GPT-5 backbone, Earth-Agent-Pro achieves 66.13% LLM-as-Judge accuracy, exceeding ReAct by 20.95 points in this metric and 24.44 points in Tools-In-Order. Joint adapter tuning raises Qwen3.5-9B LLM-as-Judge accuracy from 38.31% to 50.00%, an 11.69-point gain over the untuned configuration. Planning-only evaluation and execution with the reference workflow show that the adapters improve workflow composition and argument grounding, respectively. Code and datasets will be released soon.

I. INTRODUCTION

Earth-Agent-Pro and Earth-Bench-Pro target full-chain EO execution, where agents must find and process runtime data and ground open-ended answers in execution evidence. The benchmark holds scientific objectives fixed while moving runtime discovery and answer generation into evaluation.

A. Earth Observation Benchmarks and Datasets

  • Prior EO benchmarks and agents cover parts of the request-to-answer chain, but their tasks, input assumptions, and answer formats differ.Earth-Bench-Pro addresses this evaluation gap with matched regimes that progressively remove reference workflows and prepared inputs.Source: §I · p. 1

I. INTRODUCTION

  • Earth-Bench-Pro moves runtime data discovery, acquisition, and open-ended answer generation into evaluation while preserving Earth-Bench’s 248 scientific objectives.Its 248 Open-World Execution questions span RGB imagery, spectral observations, and remote sensing products, and require open-ended answers grounded in execution evidence.Source: §I · pp. 1–2
  • Open-world EO execution begins with a request specifying phenomenon, region, and period, then requires runtime inputs, data preparation, domain computation, and an answer grounded in evidence.Because later operations depend on accepted upstream evidence, agents must track dependencies and revise affected future operations when evidence invalidates the plan.Source: §I · pp. 1–2
  • Earth-Agent-Pro uses expert-authored skills to constrain workflow construction and runtime tool use, while structured memory links operations to accepted evidence and dependencies for localized suffix repair.Separate Planner and Executor adapters are trained for workflow composition and runtime argument grounding.Source: §I · p. 2
  • 66.13% LLM-as-Judge accuracy on Earth-Bench-OW with GPT-5 exceeds ReAct by 20.95 points; Earth-Agent-Pro also leads ReAct by 24.44 points on Tools-In-Order.Combining Planner and Executor adapters raises Qwen3.5-9B LLM-as-Judge accuracy by 11.69 points; separate evaluations assess workflow composition and argument grounding.Source: §I · p. 2

A. Earth Observation Benchmarks and Datasets

  • Earth-Bench-Pro combines 248 Open-World Execution questions with 496 matched Instruction-Following and Autonomous-Planning questions, yielding more than 700 evaluation questions.Its task composition spans application domains and task families, with global spatial distribution and temporal coverage represented in Figure 3.Source: §I · p. 2, Figure 3 · p. 3

B. Earth Observation Methods

Prior EO methods and benchmarks address perception, tool use, or parts of the analysis workflow, but do not jointly evaluate runtime data discovery through open-ended answers. Earth-Bench-Pro defines matched regimes and executable evidence chains for that evaluation.

B. Earth Observation Methods

  • EO perception models generally interpret supplied imagery or products, while agent systems extend toward tool use and workflows but rarely constrain both planning and execution with expert skills or maintain evolving evidence dependencies.The related-work review identifies role-specific alignment of workflow composition and node-level argument grounding as underexplored.Source: B. Earth Observation Methods · p. 3

A. Benchmark Overview

  • Earth-Bench-Pro has three matched regimes: Instruction-Following supplies an ordered tool sequence and prepared inputs, Autonomous-Planning hides the sequence, and Open-World Execution requires runtime data discovery or acquisition.Open-World Execution also requires open-ended answers grounded in execution evidence.Source: A. Benchmark Overview · pp. 3–4, B. Problem Formulation · p. 4
  • Earth-Bench-Pro spans eight application domains, six analysis types, regions across six inhabited continents, and requested observation periods from 1980 to 2025.Figure 3 presents task composition by domain and task family alongside global spatial distribution and temporal coverage.Source: A. Benchmark Overview · p. 4
  • Among the compared benchmarks, Earth-Bench-Pro alone covers all seven end-to-end dimensions, evaluating both final answers and the execution processes that produce them.Other benchmarks cover only parts of the workflow, such as acquisition, preprocessing, open-ended answers, or planning.Source: A. Benchmark Overview · p. 4

B. Problem Formulation

  • Each benchmark sample links a scientific query, input context or runtime data requirements, reference trajectory, and reference answer; Open-World Execution supplies runtime requirements instead of prepared inputs.In this regime, tool choice and arguments depend on the query, runtime requirements, and preceding accepted execution trace, and the answer is generated from the completed trajectory.Source: B. Problem Formulation · p. 4

C. Data Curation

Earth-Bench-Pro reconstructs and verifies executable evidence chains for 248 open-world EO tasks, then derives matched evaluation questions from those task cores.

C. Data Curation

  • Experts re-execute and correct 248 scientific task cores, replacing fixed local inputs with online acquisition and required preprocessing for 188 Product and Spectrum tasks while retaining user-provided visual inputs for 60 RGB tasks.The Product and Spectrum tasks use 323 task-level acquisition specifications covering source, product, bands, spatial and temporal extent, resolution, sampling interval, and scaling.Source: C. Data Curation · p. 5
  • Each reconstructed task pairs an open-ended request with reference evidence regenerated through execution, including recorded tool calls, returned observations, and an answer recomputed from outputs.The representation preserves artifact provenance, data dependencies, and answer evidence along the execution path.Source: C. Data Curation · p. 5
  • Quality control checks queries against input requirements, dependencies across calls, and answers against execution evidence; the 248 verified workflows contain 1,765 calls averaging 7.1 per task.Matched Instruction-Following and Autonomous-Planning variants yield 744 questions and 5,295 reference tool calls.Source: C. Data Curation · p. 5

IV. EARTH-AGENT-PRO: AN EXECUTION-ADAPTIVE AGENT FRAMEWORK

Earth-Agent-Pro separates workflow planning from runtime operation grounding and uses expert skills to constrain both stages. Its execution memory supports evidence-grounded execution and localized repair while retaining successful work.

A. Framework Overview

  • Earth-Agent-Pro separates planning an ordered workflow skeleton from grounding and executing each operation as data and intermediate artifacts emerge.The workflow and accepted execution memory are shared across roles throughout execution.Source: A. Framework Overview · p. 5
  • When a node fails or evidence remains incomplete, the Planner proposes a replacement suffix and the Executor installs it while preserving the successful prefix.The Executor exposes only verified results to subsequent operations.Source: A. Framework Overview · p. 6

B. Skill-Guided Workflow Planning

  • Six expert-authored skills specify tool scope, operation order, argument constraints, and repetition rules across acquisition, preprocessing, computation, and aggregation.The same guidance constrains workflow generation and node realization during execution.Source: B. Skill-Guided Workflow Planning · p. 6
  • A Skill Router selects a task skill and filters inapplicable tools; a Tool Router retrieves task-relevant candidates for the Planner to sequence.The full catalog contains 112 callable tools and their schemas.Source: B. Skill-Guided Workflow Planning · p. 6
  • A four-role sequential debate refines candidate tool sequences, then a catalog validator checks tool existence and active-skill membership before accepting a plan.During execution, arguments draw on the skill, tool schema, query constraints, and accepted upstream results; a checker uses the same skill when deterministic rules cannot resolve a result.Source: B. Skill-Guided Workflow Planning · p. 6

C. Workflow-Centered Structured Memory

Earth-Agent-Pro represents execution as a workflow paired with verified evidence, and uses local checks and suffix repair to address failed operations or incomplete evidence.

C. Workflow-Centered Structured Memory

  • At each step, the structured state pairs a dynamic workflow with chronologically ordered execution evidence records.Each evidence record stores the invoked tool, arguments, observation, and provenance.Source: C. Workflow-Centered Structured Memory · p. 6
  • Each workflow node tracks planned and invoked tools, status, grounded arguments, result summary, failure diagnosis, and revision; memory accumulates only verified calls.Artifact provenance records variables, time, region, bands, and batches so later nodes can resolve files, numerical series, and cross-input correspondence.Source: C. Workflow-Centered Structured Memory · p. 6

D. Evidence-Grounded Execution and Workflow Repair

  • The hybrid verification gate applies deterministic checks first, then uses a language-model checker for cases those rules cannot resolve.Retry or replace leaves persistent state unchanged; accept commits the node result and evidence record.Source: D. Evidence-Grounded Execution and… · p. 7
  • When a candidate repair passes validation, the agent preserves the successful prefix, archives the old suffix, and re-executes from the failed position onward.Repair leaves execution memory unchanged; the Executor validates candidates against the active skill and tool catalog and rejects unproductive or excessive repetitions.Source: D. Evidence-Grounded Execution and… · p. 7
  • The Answer Controller checks evidence completeness and issues an actionable diagnosis when required computations or coverage are missing.With sufficient evidence, it generates the answer only from execution memory; otherwise the Executor reopens the specified node or defaults to the final node.Source: D. Evidence-Grounded Execution and… · p. 7

V. ROLE-SPECIALIZED TRAINING

Role-specialized training separates workflow composition from runtime argument grounding, matching each adapter’s supervision to its prediction interface.

V. ROLE-SPECIALIZED TRAINING

  • The Planner composes an ordered tool sequence, while the Executor grounds a prescribed node using accepted runtime evidence.The paper argues that a monolithic trajectory objective would mix workflow errors with argument errors and spread credit across long tool chains.Source: A. Training Overview · p. 7
  • The training pipeline expands instances through dependency-aware trajectory instantiation, then extracts role-specific supervision from complete trajectories.Planner and Executor records provide separate training targets for the two interfaces.Source: A. Training Overview · p. 7
  • Planner SFT trains workflow composition, while Executor GRPO trains runtime argument grounding using group-relative advantages and locally defined objective terms.Separate adapters are optimized for the two roles.Source: A. Training Overview · p. 7

B. Dependency-Preserving Training Data Construction

The training-data construction preserves executable workflow structure while varying dependent sample fields, and Planner SFT learns tool selection and ordering from reference sequences.

B. Dependency-Preserving Training Data Construction

  • Each seed sample is decomposed into an immutable execution skeleton and a mutable instance state containing replaceable observations and sample details.The skeleton retains the reference tool sequence, schemas, and dependency structure.Source: B. Dependency-Preserving Training Data… · p. 7
  • A sample-specific program propagates root-slot replacements through dependencies while preserving the original execution skeleton.Coupled fields, such as a place and its bounding box, change together, and changed node outputs propagate to downstream arguments and observations.Source: B. Dependency-Preserving Training Data… · pp. 7–8
  • The process yields 2,480 instantiated trajectories from 248 seed workflows, then balances them with OpenEarthAgent training examples for the final trajectory set.Human reviewers inspect the slot specifications and programs before batch synthesis; the resulting data spans RGB imagery, spectral observations, and remote sensing products.Source: B. Dependency-Preserving Training Data… · p. 8

C. Planner SFT for Workflow Composition

  • Planner SFT supervises an assistant response containing only the reference tool sequence, masking prompt tokens to focus learning on tool selection and ordering.The context supplies the query, input, skill guidance, and candidate tools; the pretrained backbone is frozen while a Planner-specific LoRA module is updated.Source: C. Planner SFT for Workflow Composition · p. 8

D. Executor GRPO for Runtime Argument Grounding

Executor GRPO trains node-level argument grounding with a locally verifiable reward, while the experiments evaluate the agent across benchmarks, agent comparisons, and framework ablations.

D. Executor GRPO for Runtime Argument Grounding

  • The Executor’s node-level interface fixes the tool identity and evaluates generated arguments, isolating argument grounding from workflow decisions at other nodes.The reward compares parsed responses with reference argument dictionaries and does not require a learned reward model.Source: D. Executor GRPO for Runtime Argument… · p. 8
  • The argument reward combines key coverage, a symmetric-difference penalty, and type-aware value agreement; missing keys incur a larger penalty than extra keys.Runtime verification separately determines execution validity.Source: D. Executor GRPO for Runtime Argument… · p. 9
  • GRPO samples four candidate argument dictionaries per node and standardizes rewards within the group, updating only the Executor LoRA adapter.A KL regularizer and entropy bonus are included; explicit runtime mechanisms handle interaction, verification, retries, and workflow repair.Source: D. Executor GRPO for Runtime Argument… · p. 9

VI. EXPERIMENTS

  • The experiments assess backbone sensitivity, transfer, controlled framework comparisons, skill ablations, and the separate contributions of Planner SFT and Executor GRPO.Evaluations include OpenEarthAgent, remote-sensing MLLMs, and general-purpose agents.Source: §VI · p. 9

A. Experimental Setup

  • Earth-Bench-Pro evaluation covers end-to-end acquisition, preprocessing, analysis, and evidence-grounded answering, with final-answer and execution-trajectory metrics.The protocol includes LLM-as-Judge accuracy, ROUGE-L, semantic matching, and trajectory measures of tool coverage, ordering, exact match, and argument correctness.Source: A. Experimental Setup · p. 9
  • Framework comparisons use a shared GPT-5 backbone, and Earth-Bench-Pro numerical predictions count as correct below 15% relative error.The setup also compares with ReAct, AFlow, OpenEarthAgent, Codex, and Manus across the specified benchmarks and task categories.Source: A. Experimental Setup · p. 9

B. Main Results on Earth-Bench-Pro

Earth-Bench-Pro results show that answer quality and workflow fidelity vary across model configurations, while Earth-Agent-Pro also leads several controlled comparisons with existing agents and remote-sensing models.

B. Main Results on Earth-Bench-Pro

  • GPT-5 achieves the highest LLM-as-Judge score at 66.13%, while the fine-tuned 9B model leads TAO, TIO, TEM, and parameter accuracy, and the fine-tuned 4B model leads ROUGE-L and EM.These results show that final-answer quality and workflow fidelity do not necessarily improve together.Source: B. Main Results on Earth-Bench-Pro · p. 10, Table II · p. 10
  • Role-specialized training improves small-backbone results: Qwen3.5-9B LLM-as-Judge accuracy rises from 38.31% to 50.00%, and the 4B model gains 10.89 points in Judge accuracy and 10.97 points in TEM.The fine-tuned 9B model also raises parameter accuracy by 7.72 points.Source: B. Main Results on Earth-Bench-Pro · p. 10

C. Comparisons with Existing Methods and Agents

  • Remote-sensing MLLMs perform moderately on boundary detection and comparative ranking but drop sharply on coverage, area, and distance tasks requiring exact numerical answers.Earth-Agent-Pro achieves the highest average accuracy, with a pronounced advantage on area measurement, while retaining strong semantic-task performance.Source: C. Comparisons with Existing Methods… · p. 10
  • Earth-Agent-Pro reaches 50.58% average accuracy with Qwen3.5-9B and 50.41% with GPT-5, versus 42.60% for TerraScope, with its largest advantage in area measurement.Both Earth-Agent-Pro variants achieve 31.58% on area measurement, compared with at most 4.65% among remote-sensing MLLMs.Source: C. Comparisons with Existing Methods… · p. 10, Table III · p. 11
  • On OpenEarthAgent, Earth-Agent-Pro with GPT-5 achieves the strongest Perception, GIS, and all three tool-order scores, and improves answering accuracy by 11.16 points over the strongest GPT-5 baseline.The fine-tuned 9B model leads GIS and AnyOrder, while the 4B model leads SameOrder and Unique.Source: C. Comparisons with Existing Methods… · p. 11, Table IV · p. 11
  • Earth-Agent-Pro with GPT-5 achieves 75.00% average accuracy, 28.33 points above Codex, while reducing average latency from 268.73 s to 208.64 s.It completes all evaluated RGB examples.Source: C. Comparisons with Existing Methods… · p. 11
  • Earth-Agent-Pro leads seven of eight framework-comparison metrics, improving LLM-as-Judge accuracy by 20.95 points over the strongest baseline and raising TAO from 70.47% to 87.61%.Gains in TIO, TEM, and parameter accuracy extend beyond tool-set selection to preserving expert workflow structure.Source: C. Comparisons with Existing Methods… · p. 11

D. Ablation Studies

The ablations find complementary benefits from domain skills and role-specialized adapters, and isolate gains in workflow composition and runtime argument grounding.

Domain-Skill Ablation

  • With skills, the base Earth-Agent-Pro model raises TIO from 61.94% to 74.57% and TEM from 51.71% to 67.29%.For the fine-tuned model, adding skills improves LLM-as-Judge accuracy by 12.90 points and raises TAO, TIO, TEM, and parameter accuracy by 8.87, 10.18, 15.09, and 2.77 points, respectively.Source: Table VII · p. 12, D. Ablation Studies · p. 12

Role-Specialized Training Ablation

  • Planner and Executor tuning provide complementary benefits: Planner tuning yields larger workflow-metric gains, while Executor tuning primarily improves final-answer quality and parameter grounding.Their combination achieves the strongest overall Judge, EM, TAO, TIO, TEM, and parameter scores.Source: D. Ablation Studies · pp. 12–13, Table VIII · p. 13

Planning-Only Evaluation

  • Planner SFT reaches 92.51% TAO, 83.35% TIO, and 76.14% TEM, improving over the base Planner by 5.81, 11.48, and 9.18 points, respectively.It also exceeds GPT-5 by 3.65, 6.97, and 6.11 points on those metrics.Source: Table IX · p. 13, D. Ablation Studies · p. 13

Oracle-Plan Executor Evaluation

  • With the reference workflow fixed, Executor GRPO raises LLM-as-Judge accuracy from 46.37% to 52.82% and parameter accuracy from 34.19% to 37.20%.ROUGE-L rises from 21.92% to 25.92%, EM from 20.16% to 23.39%, and efficiency from 0.9081 to 0.9274; GPT-5 remains higher on Judge and parameter accuracy.Source: Table X · p. 13, D. Ablation Studies · p. 13

Role-Specialized Training Ablation

  • The tuned-Planner/base-Executor configuration has higher parameter accuracy than the base-Planner/tuned-Executor configuration because the end-to-end metric also depends on workflow alignment and execution progress.This motivates the isolated planning-only and oracle-plan evaluations.Source: D. Ablation Studies · p. 13

Planning-Only Evaluation

  • Planning-only scores can exceed end-to-end trajectory scores when execution failures shorten the accepted trace.Planning-only retains proposed tools without execution, whereas the complete system records only verified tool calls.Source: D. Ablation Studies · p. 13

Role-Specialized Training Ablation

  • The combined isolated evaluations attribute improved workflow composition to Planner SFT and improved argument grounding under a fixed workflow to Executor GRPO.The combined system benefits from these distinct interface-specific gains.Source: D. Ablation Studies · p. 13

VII. DISCUSSION

The representative execution cases show that quantitative EO answers depend on complete workflows, correct intermediate artifacts, and downstream operations grounded in both evidence and the scientific objective.

A. Qualitative Case Studies

  • In Spectrum UHII estimation, Earth-Agent-Pro preserves the dependency chain and recovers the reference 43.3131 K, while GPT-5 skips band extraction and Qwen-9B misgrounds thermal processing and aggregation.The comparison illustrates why plausible tools alone are insufficient when their inputs and arguments are not grounded in preceding observations.Source: A. Qualitative Case Studies · p. 14, Figure 7 · p. 14
  • For tennis-court area, Earth-Agent-Pro inserts SAM2 segmentation and obtains 879.3746 m2, consistent with the approximately 880 m2 reference, whereas GPT-5 and Qwen-9B calculate 1584.74 m2 from coarse detections.The difference depends on constructing and passing the pixel-level mask before measurement.Source: A. Qualitative Case Studies · p. 14, Figure 7 · p. 14
  • In the Product case, Earth-Agent-Pro selects incorrect downstream operations and produces 5.49%, despite correct upstream MODIS preparation; GPT-5 obtains 72.17%, near the 72.1% reference.Correct data preparation alone does not ensure that a downstream operation matches the scientific meaning of the task.Source: A. Qualitative Case Studies · p. 14, Figure 7 · p. 14

VII. DISCUSSION

  • Across the three cases, reliable answers require complete workflows, intended intermediate representations, and downstream operations grounded in accepted execution evidence and the scientific objective.The Product failure also shows that skill guidance does not fully eliminate semantic tool-selection errors.Source: A. Qualitative Case Studies · p. 14

B. Limitations and Trade-offs

The reported results pair strong benchmark performance with limits in spatial measurement, cross-modality consistency, and execution efficiency. Skills and framework design reduce logged operational errors, while the conclusion identifies argument grounding, trajectory overhead, and compact-model evidence synthesis as remaining constraints.

B. Limitations and Trade-offs

  • Zero and near-zero Distance accuracy across most methods leaves precise open-ended spatial measurement unresolved, alongside gaps in exact workflow fidelity and numerical grounding.The strongest TEM and parameter scores are 71.01% and 32.77%, respectively.Source: B. Limitations and Trade-offs · p. 15
  • Capability profiles vary by model and modality: Codex reaches 75.00% Product-task accuracy, versus 55.00% for the strongest Earth-Agent-Pro variant.Fine-tuned 4B and 9B models have higher tool-order fidelity than Earth-Agent-Pro (GPT-5) but lower answering accuracy; AFlow leads Operation F1 and ReAct leads image-generation accuracy.Source: B. Limitations and Trade-offs · p. 15
  • Fine-tuning the 9B variant raises average accuracy from 50.00% to 58.33%, while latency rises from 67.83 s to 111.79 s.Adding skills moves Efficiency farther from 1 in all three ablations, accompanying improvements in answer quality and trajectory fidelity.Source: B. Limitations and Trade-offs · p. 15

C. Error Analysis

  • Earth-Agent-Pro lowers total logged error events versus ReAct by 88.6% under NS and 94.8% under WS; skills reduce errors by 36.0% for the fine-tuned model and 60.9% for the base model.Figure 8 counts logged operational errors; a trajectory can contribute multiple events.Source: Figure 8 · p. 15, C. Error Analysis · p. 15

VIII. CONCLUSION

  • Earth-Agent-Pro leads seven of eight framework-comparison metrics under a shared GPT-5 backbone and reaches 66.13% LLM-as-Judge accuracy.The conclusion also reports higher workflow-composition scores for the tuned Planner and higher answer and parameter scores for the tuned Executor under a fixed workflow.Source: §VIII · p. 15
  • Remaining constraints include argument grounding under runtime uncertainty, trajectory overhead from skill-guided execution, and evidence synthesis in compact models.These are identified as priorities for future work.Source: §VIII · p. 15

Related papers

Loading 2609.12533v1…