VRoPE: Rotary Position Embedding for Video Large Language Models

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Liu, Zikang, Guo, Longteng, Tang, Yepeng, Yue, Tongtian, Cai, Junxian, Ma, Kai, Liu, Qingbin, Chen, Xi, Liu, Jing
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866912678683869184
author Liu, Zikang
Guo, Longteng
Tang, Yepeng
Yue, Tongtian
Cai, Junxian
Ma, Kai
Liu, Qingbin
Chen, Xi
Liu, Jing
author_facet Liu, Zikang
Guo, Longteng
Tang, Yepeng
Yue, Tongtian
Cai, Junxian
Ma, Kai
Liu, Qingbin
Chen, Xi
Liu, Jing
contents Rotary Position Embedding (RoPE) has shown strong performance in text-based Large Language Models (LLMs), but extending it to video remains a challenge due to the intricate spatiotemporal structure of video frames. Existing adaptations, such as RoPE-3D, attempt to encode spatial and temporal dimensions separately but suffer from two major limitations: positional bias in attention distribution and disruptions in video-text transitions. To overcome these issues, we propose Video Rotary Position Embedding (VRoPE), a novel positional encoding method tailored for Video-LLMs. Specifically, we introduce a more balanced encoding strategy that mitigates attention biases, ensuring a more uniform distribution of spatial focus. Additionally, our approach restructures positional indices to ensure a smooth transition between video and text tokens. Extensive experiments on different models demonstrate that VRoPE consistently outperforms previous RoPE variants, achieving significant improvements in video understanding, temporal reasoning, and retrieval tasks. Code is available at https://github.com/johncaged/VRoPE.
format Preprint
id arxiv_https___arxiv_org_abs_2502_11664
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle VRoPE: Rotary Position Embedding for Video Large Language Models
Liu, Zikang
Guo, Longteng
Tang, Yepeng
Yue, Tongtian
Cai, Junxian
Ma, Kai
Liu, Qingbin
Chen, Xi
Liu, Jing
Artificial Intelligence
Rotary Position Embedding (RoPE) has shown strong performance in text-based Large Language Models (LLMs), but extending it to video remains a challenge due to the intricate spatiotemporal structure of video frames. Existing adaptations, such as RoPE-3D, attempt to encode spatial and temporal dimensions separately but suffer from two major limitations: positional bias in attention distribution and disruptions in video-text transitions. To overcome these issues, we propose Video Rotary Position Embedding (VRoPE), a novel positional encoding method tailored for Video-LLMs. Specifically, we introduce a more balanced encoding strategy that mitigates attention biases, ensuring a more uniform distribution of spatial focus. Additionally, our approach restructures positional indices to ensure a smooth transition between video and text tokens. Extensive experiments on different models demonstrate that VRoPE consistently outperforms previous RoPE variants, achieving significant improvements in video understanding, temporal reasoning, and retrieval tasks. Code is available at https://github.com/johncaged/VRoPE.
title VRoPE: Rotary Position Embedding for Video Large Language Models
topic Artificial Intelligence
url https://arxiv.org/abs/2502.11664