<p>Open-Vocabulary Object Detection (OVD) aims to detect and localize novel object categories using a detector trained only on base categories. Current mainstream approaches typically transfer image-text alignment capabilities from large pre-trained models to enable novel object generalization. While these methods have achieved impressive performance, they often exhibit limited gains in localization accuracy. To address this limitation, Segment Anything Model (SAM) offers a promising solution thanks to its precise boundary-aware segmentation capability. However, SAM produces category-agnostic masks that are not directly compatible with object detection due to a fundamental task misalignment: SAM segments regions without semantic labels, whereas OVD requires class-specific bounding boxes. To bridge this gap, we propose a reinforcement learning-based model collaboration framework that effectively integrates the precise boundary segmentation of the SAM into OVD models. The core of our framework is model collaboration agent (MCA), which is a spatial- and semantic-aware GRPO-based module to dynamically select semantically relevant SAM-generated regions as visual prompts for the OVD model. Our proposed MCA leverages feedback signals from the OVD model to iteratively refine its selection, transforming raw segmentation masks into high-quality, context-aware visual prompts that effectively guide the detector toward more accurate localization. Extensive experiments on COCO, LVIS, and ODinW-35 demonstrate significant improvements in both classification and localization performance, achieving state-of-the-art results and validating the effectiveness of our method.</p>