Earth-Bench-V2

Earth-Bench-V2 evaluates Earth observation agents on 248 task cores with prepared observations, reference tool trajectories, and final-answer annotations. Tasks cover spectral analysis, remote sensing products, and RGB perception.

This release continues the Earth-Bench benchmark introduced in Earth-Agent. Its annotation file includes both Instruction Following (IF) and Autonomous Planning (AP) evaluation entries for each task. Earth-Agent-Pro studies these prepared-input regimes alongside Open-World Execution (OW). The companion OW release is Earth-Bench-OW.

The local release name is Earth-Bench-V2; the Pro paper calls the full three-regime benchmark Earth-Bench-Pro. This repository contains the prepared-input portion, with 248 task records and 496 IF/AP evaluation entries.

Papers and code

• Earth-Agent: Unlocking the Full Landscape of Earth Observation with Agents (ICLR 2026).
• Earth-Agent-Pro: Towards Real-World Full-Chain Earth Observation with Agents.
• Earth-Agent code and evaluation framework.

Task coverage

Modality	Task IDs	Task cores
Spectrum	1–100	100
Products	101–188	88
RGB	189–248	60
Total	1–248	248

Task IDs match the companion release. The number of local files reflects this snapshot; it is not the number of questions or evaluation instances.

Updated benchmark results

The following measurements use the updated Earth-Bench-V2 benchmark.

Modality	Model	Acc (%)	Eff	TAO (%)	TIO (%)	TEM (%)	Param (%)
Spectrum	GPT-5	59.00	2.0121	66.65	61.64	32.39	21.61
Spectrum	Qwen3-Max	50.00	2.9887	70.10	57.14	8.02	11.36
Products	GPT-5	83.54	1.4616	57.78	45.60	32.72	18.64
Products	Qwen3-Max	68.18	1.4877	70.10	73.85	38.62	25.89
RGB	GPT-5	75.00	1.8253	82.89	77.82	48.13	42.21
RGB	Qwen3-Max	38.33	0.8607	59.67	38.33	35.25	36.67
All	GPT-5	71.13	1.7716	67.43	59.86	36.31	25.54
All	Qwen3-Max	53.69	1.9370	68.90	54.85	25.54	21.20

Acc: Accuracy; Eff: Efficiency; TAO: Tool-Any-Order; TIO: Tool-In-Order; TEM: Tool-Exact-Match; Param: Parameter Accuracy.

Repository layout

Earth-Bench-V2/
├── README.md
├── LICENSE
├── THIRD_PARTY_NOTICES.md
├── task_index.jsonl
├── question.json
├── data/question1/ ... data/question248/
├── model_results.csv
├── model_out/
└── tools/

Component	Contents
question.json	A JSON object keyed by task IDs "1" through "248"; 248 records.
data/	13,523 prepared input files, organized by task ID.
dialogs	User request, reference assistant/tool interactions, and final answer.
choices	Candidate answers for 247 tasks; task 163 has null and a numeric reference answer.
evaluation	Two entries per task: IF and AP; includes question, ground-truth whitelist/blacklist, metric, and regime.
tools, files	Reserved fields; both are null in this release.

model_results.csv is a semicolon-delimited expert-model response table with columns file_path, model, text_prompt, bbox, and result. model_out/ holds 26 stored output masks used by the perception tools. These files support replay and are not leaderboard scores.

task_index.jsonl provides one derived row per task for browsing: question_id, modality, question, reference_answer, and choices. The original JSON file remains the source of truth for trajectories, acquisition specifications, and evaluation details. This release provides a single test index and no official training split.

Download

Install huggingface_hub, then download from the Hub. The full snapshot occupies approximately 15.1 GB.

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="LvZTTTT/Earth-Bench-V2",
    repo_type="dataset",
    local_dir="benchmark",
)

For annotations only:

snapshot_download(
    repo_id="LvZTTTT/Earth-Bench-V2",
    repo_type="dataset",
    local_dir="benchmark",
    allow_patterns=["README.md", "question.json", "task_index.jsonl"],
)

Read the annotations

import json
from pathlib import Path

questions = json.loads(Path("benchmark/question.json").read_text())
example = questions["1"]
request = example["dialogs"][0]["content"]
reference_trajectory = example["dialogs"][1:-1]
reference_answer = example["dialogs"][-1]["content"]
evaluation = example["evaluation"]

Launch tools from a working directory where benchmark/ resolves to the downloaded repository. The bundled MCP configuration uses tools/... paths; either install those scripts under your framework's tools/ directory or update the configuration to their actual paths. Configure the output directory consistently with the reference benchmark/out/... paths.

Evaluation and reproducibility

Use the papers for the regime definitions, final-answer evaluation, and trajectory metrics. The original Earth-Agent framework provides evaluation code; integration of this snapshot requires selecting the correct annotation file, regime, data paths, and tool implementation. This dataset repository does not include an end-to-end evaluation runner.

• Use prepared inputs for IF/AP. For IF, expose the prescribed tool sequence according to the evaluation protocol; for AP, hide that sequence and let the agent plan. Never expose reference answers or reference tool outputs to the evaluated agent.
• Ground-truth regime labels vary in capitalization; normalize them when parsing.
• Task 163 requests 2022-09-12, but its reference input filenames contain 2020-09-12. The original annotation is preserved. Record any correction or exclusion when reporting scores.

Intended use

Use this benchmark to evaluate scientific Earth observation workflows, tool selection, argument grounding, and final-answer quality. The released trajectories and answers are evaluation references. Disclose any use of them for training or prompt construction before reporting benchmark performance.

License and sources

Benchmark annotations and release documentation are licensed under CC BY-NC 4.0. Attribute the benchmark and cite the relevant papers. Third-party satellite products, RGB imagery, and inherited code retain their original terms; see THIRD_PARTY_NOTICES.md. The benchmark license does not replace those terms.

Citation

Please cite both papers when using these matched benchmark releases.

@article{feng2025earthagent,
  title   = {Earth-Agent: Unlocking the Full Landscape of Earth Observation with Agents},
  author  = {Feng, Peilin and Lv, Zhutao and Ye, Junyan and Wang, Xiaolei and Huo, Xinjie and Yu, Jinhua and Xu, Wanghan and Zhang, Wenlong and Bai, Lei and He, Conghui and Li, Weijia},
  journal = {arXiv preprint arXiv:2509.23141},
  year    = {2025},
  url     = {https://arxiv.org/abs/2509.23141}
}

@article{lv2026earthagentpro,
  title   = {Earth-Agent-Pro: Towards Real-World Full-Chain Earth Observation with Agents},
  author  = {Lv, Zhutao and Dang, Chenhao and Feng, Yi and Gong, Yanpei and Wang, Xiaolei and Ye, Junyan and He, Conghui and Li, Weijia},
  journal = {arXiv preprint arXiv:2609.12533},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.12533}
}