VRoPE: Rotary Position Embedding for Video Large Language Models

Official implementation of VRoPE: Rotary Position Embedding for Video Large Language Models

Resources: [📑 Paper (Arxiv)]

News

• [2025/11/18] We have updated our official implementation of VRoPE.
• [2025/10/31] Our paper has been accepted by EMNLP 2025 (Main)! We are cleaning up the code for release and will publish it soon—stay tuned!
• [2025/5/21] We updated our paper on arxiv.
• [2025/2/18] Paper is now available on arxiv.

Results

• Results on Video-NIAH

Todo

• [✔] Release Paper
• [✔] Release Code (After the paper acceptance)

Citation

If you find this paper helpful, you can cite the following paper:

@misc{liu2025vroperotarypositionembedding,
    title={VRoPE: Rotary Position Embedding for Video Large Language Models}, 
    author={Zikang Liu and Longteng Guo and Yepeng Tang and Tongtian Yue and Junxian Cai and Kai Ma and Qingbin Liu and Xi Chen and Jing Liu},
    year={2025},
    eprint={2502.11664},
    archivePrefix={arXiv},
    primaryClass={cs.AI},
    url={https://arxiv.org/abs/2502.11664}, 
}