Geo-R1: Reinforcement Fine-Tuning for Few-Shot Geospatial Referring

Geo-R1 is a paradigm for reasoning-centric reinforcement fine-tuning (RFT) aimed at enhancing few-shot understanding of geospatial referring expressions in remote sensing. It tackles the challenges faced by supervised fine-tuning (SFT) on multimodal large language models, particularly in situations with scarce data. The approach requires the model to initially create clear, interpretable reasoning chains that break down referring expressions, subsequently using these rationales to identify target objects. This method of prioritizing reasoning before action boosts both generalization and interpretability, even with minimal annotations. The effectiveness of the model has been assessed using three benchmarks for few-shot geospatial referring.