TVR-Qwen3.5-9B-VA-SFT-RL

Best-performing model from the paper "Where to Look: Can Foundation Models Reach a Target Viewpoint Through Active Exploration?" [arXiv].

Model Description

Fine-tuned from Qwen3.5-9B for Target Viewpoint Reproduction (TVR) — the agent receives a target image and navigates in a 3D indoor environment to reproduce the exact viewpoint through active exploration.

• Base model: Qwen3.5-9B (8.95B parameters, full fine-tuning)
• Training: Visual-Action SFT + Online Multi-turn GRPO
• Overall success rate: 51.4% on TVRBench (500 tasks)
• Image resolution: 640×360 (~220 visual tokens per image)
• Action space: 9 actions (MoveAhead/Back/Left/Right, RotateLeft/Right, LookUp/Down, Stop)

Usage

See the TVRBench repository for evaluation instructions.

Citation

@misc{li2026lookfoundationmodelsreach,
      title={Where to Look: Can Foundation Models Reach a Target Viewpoint Through Active Exploration?}, 
      author={Liyang Li and Muzhi Zhu and Zhiyue Zhao and Hengyu Zhao and Ke Liu and Linhao Zhong and Hao Chen and Chunhua Shen},
      year={2026},
      eprint={2606.01247},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.01247}, 
}