VRoPE: Rotary Position Embedding for Video Large Language Models
This work solves the problem of effective positional encoding for video-LLMs, enabling better spatiotemporal modeling, though it is incremental as it builds on existing RoPE methods.
The paper tackled the challenge of extending Rotary Position Embedding (RoPE) to video large language models by addressing positional bias and video-text transition disruptions, resulting in VRoPE, which outperforms previous variants with significant improvements in video understanding, temporal reasoning, and retrieval tasks.
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.