VRoPE: Rotary Position Embedding for Video Large Language Models

TL;DR

VRoPE introduces a novel rotary position embedding tailored for Video-LLMs, effectively addressing spatial-temporal encoding challenges and improving video understanding and reasoning tasks.

Contribution

It presents a new balanced positional encoding method for Video-LLMs that overcomes limitations of previous adaptations like RoPE-3D.

Findings

VRoPE outperforms previous RoPE variants in video understanding tasks.

It achieves significant improvements in temporal reasoning.

The method ensures a more uniform spatial focus distribution.

Abstract

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.