Accepted by ACM Multimedia 2026 (ACM MM 2026)
Remote sensing tasks increasingly require more than one-shot perception: they involve long processing chains with multi-step reasoning, tool invocation, intermediate-result interpretation, and iterative decision-making. Existing monolithic agents can struggle with workflow organization, domain expertise, and error recovery as mistakes propagate across stages.
HiRS-Agent is a hierarchical multi-agent system designed for reliable long-horizon remote sensing task solving. It organizes execution through a two-level Manager-Specialist architecture:
- The Manager Layer performs task planning, dynamic routing, step-level verification, replanning, and termination control.
- The Specialist Layer aligns domain-specific reasoning and tools with the remote sensing workflow, spanning surface parsing, physical retrieval, and statistical analytics.
This structure enables workflow-aware organization, remote-sensing-specialized execution, and verification-guided control throughout a multi-stage task.
- Hierarchical remote sensing agents. A Manager-Specialist architecture models the stage-dependent structure of real remote sensing workflows.
- Verification-guided workflow control. Intermediate results are checked under domain constraints, enabling rerouting, replanning, and recovery.
- Expert-to-Workflow Alignment Tuning. A two-stage supervised tuning strategy first injects remote sensing expertise and then aligns task intent with executable workflows.
- Verification-Guided Hierarchical Reinforcement Learning. Layer-aware optimization jointly improves global orchestration and local tool execution.
- Evaluation across tool environments. Experiments on Earth-Bench and ThinkGeo demonstrate consistent gains in long-horizon tool use and final-task correctness.
On Earth-Bench, HiRS-Agent with a Qwen3-4B backbone improves final accuracy from 15.73/10.08 to 43.95/45.56 under the AP/IF settings. Tool Exact Match rises from 0.00/8.63 to 31.67/34.64, while execution efficiency also improves.
On ThinkGeo, the same backbone improves instruction/tool/argument scores from 18.35/8.54/1.24 to 73.73/47.87/8.51, demonstrating transfer across a different tool library without task-specific optimization on ThinkGeo.
- 2026: HiRS-Agent was accepted by ACM Multimedia 2026.
- Coming soon: Paper, code, model checkpoints, training data, and reproduction instructions.
This repository currently serves as the official project homepage. The implementation and complete reproduction guide are being prepared and will be released in stages:
- Paper and citation metadata
- Inference and evaluation code
- Training recipes for Expert-tuning and VG-HRL
- Model checkpoints
- Data preparation and benchmark instructions
- Reproducible environments and example cases
Please watch this repository for release updates.
The camera-ready BibTeX entry will be added after the complete paper metadata is publicly available.
We thank the developers and maintainers of the open-source models, remote sensing benchmarks, and tool ecosystems that support this research. Detailed acknowledgements will accompany the code release.