DeepGrounder: Generalized Reasoning for MLLM-Based Remote Sensing Visual Grounding

Remote Sensing Visual Grounding (RSVG) aims to accurately localize target regions in remote sensing imagery based on natural language descriptions. Existing approaches leveraging multimodal large language models (MLLMs) predominantly rely on supervised fine-tuning (SFT). While SFT enhances instruction-following capabilities, it often fails to foster genuine reasoning due to overlearning statistical correlations among image features, text, and coordinates. This limitation severely hampers generalization in both in-domain scenarios with imbalanced category difficulty and cross-domain scenarios with distribution shifts. To address this, we propose DeepGrounder, a novel RSVG method designed for robust generalization and reasoning. DeepGrounder employs a two-stage training strategy: the first stage uses SFT to inject domain knowledge and establish foundational spatial perception; the second stage introduces reinforcement fine-tuning (RFT) via Group Relative Policy Optimization (GRPO) to stimulate structured reasoning. To provide stable and verifiable learning signals for multimodal reasoning, we formulate a hybrid verifiable reward mechanism encompassing reasoning format standardization, geometric localization accuracy, and vision-language semantic consistency. This mechanism guides the model from passive mapping to active reasoning. Experiments on RRSIS-D and DIOR-RSVG datasets show that DeepGrounder significantly outperforms SFT-based baselines, particularly on challenging categories. Notably, its zero-shot performance on VRSBench surpasses models specifically fine-tuned on that dataset, directly validating its superior cross-domain generalization capabilities. The code will be available at https://github.com/ULTA-Web/DeepGrounder.