SimCRAFT （EMNLP 2026 Main）

Official repository for "SimCRAFT: Distilling Remote Sensing Agents via Synthetic Trajectories and Contextual Retrieval-Augmented Fine-Tuning".

SimCRAFT distills expert-level remote sensing agent capabilities into a compact 7B model through synthetic trajectory generation and contextual retrieval-augmented fine-tuning, achieving GPT-4-level planning performance with ~100× fewer parameters.

News

• Our paper has been accepted as main conference in EMNLP 2026

🚧 Release Status

Code, dataset, and model weights will be publicly released upon paper acceptance.

TODO

• SimRS-14K dataset and Procedural Knowledge Base (PKB)
• 36 atomic RS tool specifications
• Mock Execution Engine
• Multi-agent trajectory synthesis pipeline
• CRAFT training code
• SimCRAFT-7B model weights
• Evaluation scripts and reproduction instructions

Please ⭐️ the repository to stay updated.

📑 Citation

@article{wang2026simcraft,
      title={SimCRAFT: Distilling Remote Sensing Agents via Synthetic Trajectories and Contextual Retrieval-Augmented Fine-Tuning}, 
      author={Wang, Haoran and Yao, Jing and Yang, Xu and Wang, Zeqing and Zhang, Yang and Ghamisi, Pedram and Chen, Zhengchao},
      journal={arXiv preprint arXiv:2608.30277},
      year={2026},
      url={https://arxiv.org/abs/2608.30277}
}