stvg-r1-model-7b

STVG-R1: Incentivizing Instance-Level Reasoning and Grounding in Videos via Reinforcement Learning

<div align="center">

</div>

This repository contains the checkpoint for STVG-R1, the first reinforcement learning framework for spatial–temporal video grounding. We finetuned Qwen2.5-VL-7B-Instruct with constructed Visual Prompt Data.

---

📋 Key Features

• Object-Centric Visual Prompt: A simple yet effective object-centric visual prompting paradigm reformulates dense per-frame coordinate prediction into a compact object ID identification task.
• Reinforcement Learning Training: STVG-R1 is trained entirely using reinforcement learning, enhancing its ability to generate accurate spatial temporal visual grounding results.
• Strong Zero-Shot Generalization: STVG-R1 exhibits strong zero-shot generalization to multi-object referring video object segmentation task, despite being trained only on single-object grounding data.
• SOTA Performance: STVG-R1 sets a new SOTA on the HCSTVG-v1, HCSTVG-v2, ST-Align and MeViS benchmarks.

---

💾 Model Details

• Base Model: Qwen2.5-VL-7B
• Training Data: HCSTVG-v1, HCSTVG-v2 and VidSTG

---

📝 Citation

If you use STVG-R1 or this checkpoint in your research, please cite:

@article{zhang2026stvg,
  title={STVG-R1: Incentivizing Instance-Level Reasoning and Grounding in Videos via Reinforcement Learning},
  author={Zhang, Xiaowen and Gao, Zhi and Jiao, Licheng and Li, Lingling and Li, Qing},
  journal={arXiv preprint arXiv:2602.11730},
  year={2026}
}