Teaching General-Purpose Agentsto Act on EarthviaSkill Evolution

A Self-evolving Skill Framework for Tool-Driven Geospatial Intelligence

Haojun Tang1,2 · Wenda Zhao2,✉ · Xiang Zhou2 · Qiwei Lin3 · Lida Huang1 · Yan Wang1,✉

1 Institute for AI Industry Research (AIR), Tsinghua University2 Dalian University of Technology3 Beijing Institute of Radio Measurement

✉ Corresponding authors

SKILLEARTH · TOOL-DRIVEN GEOSPATIAL INTELLIGENCE

From trajectories to
evolvable Earth skills.

SkillEarth converts demonstrated remote-sensing tool trajectories into reusable skills that guide general-purpose agents to plan, execute, and improve geospatial workflows—without retraining the model.

+21.69pp · largest E2E metric gain
3.18×Qwen3 General Accuracy
7 / 7backbones improve in macro E2E
Explore the framework
CONCRETE ARTIFACT

What a skill looks like?

EXTRACTED SKILL

nbr-burn-severity-and-regrowth-change-analysis

Released SkillBank ↗
WHEN TO USE

Use when a user requests a visual analysis of burn severity and vegetation recovery changes over time using normalized burn ratio (NBR) index differences, requiring classification of low, moderate, severe burn, unburned, and enhanced regrowth areas.

PLAN
  1. Call GetAreaBoundary with the specified geographic region to define the study area and generate a geopackage file that serves as the shared workspace.
  2. Call AddIndexLayer with NBR and the first date to compute and store the baseline NBR raster layer.
  3. Call AddIndexLayer a second time with a different month to compute the later NBR raster layer.
  4. Call ComputeIndexChange with the two distinct NBR layers to classify burn severity and regrowth patterns.
  5. … Call ShowIndexLayer, then terminate only after the preview and base layers are successfully generated.
GROUNDED OBSERVATION ROUTING
  • Use the gpkg output from GetAreaBoundary to provide the gpkg argument for both AddIndexLayer calls.
  • Use the gpkg output from GetAreaBoundary to provide the gpkg argument for ComputeIndexChange and ShowIndexLayer.
  • Use the layer_name outputs from AddIndexLayer to provide layer1_name and layer2_name for ComputeIndexChange.
  • Use the diff_layer_name output from ComputeIndexChange to provide the layer_name argument for ShowIndexLayer.
  • … Use the text output from ShowIndexLayer as evidence for the ans argument of Terminate.
TOOL LIST
  • GetAreaBoundaryarea, buffer_m
  • AddIndexLayergpkg, index_type, layer_name, year, month
  • ComputeIndexChangegpkg, index_type, layer1_name, layer2_name, diff_layer_name
  • ShowIndexLayergpkg, index_type, layer_name
  • Terminateans
01 Framework

Teach agents to act
on Earth—once and again.

Geospatial tasks often require long, brittle chains of tool use. SkillEarth preserves what works in those trajectories as modular, retrievable skills that can improve future executions.

01

Skill Extraction

Extract controllable, reusable skills from the components of demonstrated tool trajectories.

→
02

Skill Evolution

Use dual-judge feedback and a GOR graph for grounded, fine-grained skill refinement.

→
03

Skill Decision

Decide whether each evolved skill should be added, merged into the bank, or discarded.

Overview of the SkillEarth framework
Figure 1. SkillEarth extracts, evolves, and retrieves skills to guide tool-driven Earth intelligence.
02 Generalization

Skill guidance generalizes
across seven LLMs.

Across three frontier and four open-source models, SkillEarth improves the macro-average of all nine OpenEarth-Bench end-to-end metrics from 29.66 to 37.36.

Macro-average E2E+7.70 pp

25.96% relative improvement across nine metrics.

Largest model gain+12.64 pp

Qwen3-4B-Instruct, a 4B open-weight model.

Coverage7 / 7

All evaluated models improve in macro E2E score.

End-to-end macro score is the unweighted mean of Per., Op., Logic., GIS., AnyOr., SameO., Uni., Ans., and Gen. Values are derived from the paper’s OpenEarth-Bench table.

TABLE · HIDDEN BY DEFAULTExplore end-to-end task performance, model by model.

All seven backbones, without/with SkillEarth macro scores, and per-model gains.

Click to expandClick to collapse
BackboneFamilyWithout skillWith SkillEarthΔ macro E2E
Each value is the macro-average across nine end-to-end OpenEarth-Bench metrics. The full per-metric results and evaluation trajectories are available in the project repository.
03 Transferability

Skills transfer across
model families.

A SkillBank is not tied to the model that extracted it. Transferred skills improve the end-to-end general accuracy of both Qwen3-4B-Instruct and DeepSeek-V4-Flash.

Best cross-model lift+15.86 pp

General accuracy for Qwen3-4B-Instruct using skills extracted by Doubao-Seed-2.0-Mini.

without skill with transferred skill

Bars report the Gen. end-to-end accuracy from the cross-model transfer experiments. The colored source cards identify the model used to construct the transferred SkillBank.

FULL BENCHMARK COMPARISONS

How does SkillEarth compare with state-of-the-art methods?

Complete results against state-of-the-art remote-sensing agents and skill-extraction methods on both OpenEarth-Bench and Earth-Bench.

OPENEARTH-BENCH · TABLE 1

Remote-sensing agents

Qwen3-4B-Instruct · EP: expert priors · FT: fine-tuning

MethodEPFTStep-by-StepEnd-to-End
Ins. ↑Tool. ↑ArgN. ↑ArgV. ↑Summ. ↑Per. ↑Op. ↑Logic. ↑GIS. ↑AnyOr. ↑SameO. ↑Uni. ↑Ans. ↑Gen. ↑
OpenEarthAgent✗✗97.3486.9484.1233.5583.2813.8620.9518.3571.8216.0014.7121.4713.7215.86
Earth-Agent✗✗99.9890.8487.3133.3788.7031.2614.6711.9451.582.482.314.1912.8219.31
RSMeM✓✗99.2087.9785.0633.0584.9231.207.7815.0853.569.418.9811.4618.1933.10
RS-Agent✓✗100.0043.6640.1711.0488.3128.5219.6718.7212.673.933.935.829.4637.93
SkillEarth✗✗99.0791.0486.4836.1792.6735.0033.0628.7381.8525.3225.1527.8919.5924.14
EARTH-BENCH · TABLE 2

Remote-sensing agents

Qwen3-4B-Instruct · EP: expert priors · FT: fine-tuning · AP and IF settings

MethodEPFTAutomatic Planning (AP)Instruction Following (IF)
Step-by-StepEnd-to-EndStep-by-StepEnd-to-End
AnyOr. ↑InOr. ↑Exact. ↑Param. ↑Eff. ↓Acc. ↑AnyOr. ↑InOr. ↑Exact. ↑Param. ↑Eff. ↓Acc. ↑
OpenEarthAgent✗✗46.5539.0128.5918.982.709918.1550.2141.6830.7520.752.777819.76
Earth-Agent✗✗20.471.320.000.003.974715.7329.6910.898.634.215.455410.08
RSMeM✓✗50.4333.257.687.511.170321.4640.7223.2713.931.160.483921.46
HiRS-Agent✓✓50.9541.1031.6719.542.505943.9553.7946.3434.6420.813.301345.56
RS-Agent✓✗22.6815.8415.178.902.338529.4427.2518.6117.749.182.229533.06
SkillEarth✗✗53.1743.2435.4222.541.865922.9860.1851.3847.6428.241.964726.61
OPENEARTH-BENCH · TABLE 3

Skill-extraction methods

Qwen3-4B-Instruct · shared evaluation protocol

MethodStep-by-StepEnd-to-End
Ins. ↑Tool. ↑ArgN. ↑ArgV. ↑Summ. ↑Per. ↑Op. ↑Logic. ↑GIS. ↑AnyOr. ↑SameO. ↑Uni. ↑Ans. ↑Gen. ↑
Trace2Skill98.8889.8985.6136.3584.4313.2130.5622.9740.332.402.403.519.6714.48
AutoSkill98.7883.6079.6138.3587.3117.9724.8820.0235.963.853.764.968.319.66
SkillX99.0390.3785.6234.1984.6815.8029.7220.1272.6217.1116.7720.8720.1217.24
SkillRevise98.5188.3084.0733.6083.8313.5025.2621.3135.502.402.223.256.6314.48
SkillEarth99.0791.0486.4836.1792.6735.0033.0628.7381.8525.3225.1527.8919.5924.14
EARTH-BENCH · TABLE 4

Skill-extraction methods

Qwen3-4B-Instruct · AP: automatic planning · IF: instruction following

MethodAutomatic Planning (AP)Instruction Following (IF)
Step-by-StepEnd-to-EndStep-by-StepEnd-to-End
AnyOr. ↑InOr. ↑Exact. ↑Param. ↑Eff. ↓Acc. ↑AnyOr. ↑InOr. ↑Exact. ↑Param. ↑Eff. ↓Acc. ↑
Trace2Skill46.1433.0326.3420.481.885420.8149.5539.3932.5324.292.053224.60
AutoSkill43.5628.6317.2113.121.671118.5548.3435.4425.5218.011.834422.18
SkillX51.5442.9533.4320.661.992220.1958.0048.8744.1727.271.977218.15
SkillRevise46.4234.6824.9319.221.760922.1847.6535.7930.0321.861.803422.58
SkillEarth53.1743.2435.4222.541.865922.9860.1851.3847.6428.241.964726.61

Top three results in each column are highlighted in red, blue, and orange.

04 Citation

Build agents that
learn to act on Earth.

Explore the paper, reproducible evaluation results, and released SkillBanks.

@misc{tang2026skillearth,
  title  = {SkillEarth: Teaching General-Purpose
            Agents to Act on Earth via Skill Evolution},
  author = {Haojun Tang and Wenda Zhao and others},
  year   = {2026},
  eprint = {2611.15481}
}