cs.CVJul 22, 2026

PCA: Persistence-Aware Compression and Aggregation for Fast Video Large Language Models

Authors: Zihan SongShuo YeBo ZhaoRuixin ZhangJiayu ZhangShouhong DingZitong Yu

Organizations: Institute for Artificial Intelligence, Great Bay University · Sun Yat-sen University · Youtu Lab, Tencent · Guangdong Provincial Key Laboratory of Intelligent Information Processing & Shenzhen Key Laboratory of Media Security, Shenzhen University · Dongguan Key Laboratory for Intelligence and Information Technology

Abstract

Despite advances in Video Large Language Models (VLLMs) that have displayed promising outcomes in video understanding, the redundancy in the long-duration frames remains a hindrance to efficient reasoning. This paper introduces a training-free P\mathbf{P}ersistence-Aware C\mathbf{C}ompression and A\mathbf{A}ggregation (PCA) method designed to preserve high-fidelity raw visual information before the encoding stage. PCA can be built on arbitrary VLLMs and consists of two modules: 1) A Dynamic Downsampling (DD) module that adaptively removes redundant frames by analyzing frame-wise similarity. 2) A Persistence-Aware Motion Enhancement (PAME) module that enriches each selected keyframe by aggregating the temporal context of its neighbors, ensuring that essential information is preserved even after aggressive frame reduction. Our approach substantially reduces the computation of long-context modeling, while enhancing the performance of the baseline model. Extensive experiments demonstrate that PCA consistently outperforms existing state-of-the-art approaches in both efficiency and accuracy, achieving a speedup of 1.8×\times to 2.5×\times compared to the baseline VLLM. The code is open-sourced at https://github.com/Heisenberg10110/PCA.

Explore similar work

CardsList