Multi-Granular Spatio-Temporal Token Merging for Training-Free Acceleration of Video LLMs

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Hyun, Jeongseok, Hwang, Sukjun, Han, Su Ho, Kim, Taeoh, Lee, Inwoong, Wee, Dongyoon, Lee, Joon-Young, Kim, Seon Joo, Shim, Minho
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908445015277568
author Hyun, Jeongseok
Hwang, Sukjun
Han, Su Ho
Kim, Taeoh
Lee, Inwoong
Wee, Dongyoon
Lee, Joon-Young
Kim, Seon Joo
Shim, Minho
author_facet Hyun, Jeongseok
Hwang, Sukjun
Han, Su Ho
Kim, Taeoh
Lee, Inwoong
Wee, Dongyoon
Lee, Joon-Young
Kim, Seon Joo
Shim, Minho
contents Video large language models (LLMs) achieve strong video understanding by leveraging a large number of spatio-temporal tokens, but suffer from quadratic computational scaling with token count. To address this, we propose a training-free spatio-temporal token merging method, named STTM. Our key insight is to exploit local spatial and temporal redundancy in video data which has been overlooked in prior work. STTM first transforms each frame into multi-granular spatial tokens using a coarse-to-fine search over a quadtree structure, then performs directed pairwise merging across the temporal dimension. This decomposed merging approach outperforms existing token reduction methods across six video QA benchmarks. Notably, STTM achieves a 2$\times$ speed-up with only a 0.5% accuracy drop under a 50% token budget, and a 3$\times$ speed-up with just a 2% drop under a 30% budget. Moreover, STTM is query-agnostic, allowing KV cache reuse across different questions for the same video. The project page is available at https://www.jshyun.me/projects/sttm.
format Preprint
id arxiv_https___arxiv_org_abs_2507_07990
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Multi-Granular Spatio-Temporal Token Merging for Training-Free Acceleration of Video LLMs
Hyun, Jeongseok
Hwang, Sukjun
Han, Su Ho
Kim, Taeoh
Lee, Inwoong
Wee, Dongyoon
Lee, Joon-Young
Kim, Seon Joo
Shim, Minho
Computer Vision and Pattern Recognition
Artificial Intelligence
Video large language models (LLMs) achieve strong video understanding by leveraging a large number of spatio-temporal tokens, but suffer from quadratic computational scaling with token count. To address this, we propose a training-free spatio-temporal token merging method, named STTM. Our key insight is to exploit local spatial and temporal redundancy in video data which has been overlooked in prior work. STTM first transforms each frame into multi-granular spatial tokens using a coarse-to-fine search over a quadtree structure, then performs directed pairwise merging across the temporal dimension. This decomposed merging approach outperforms existing token reduction methods across six video QA benchmarks. Notably, STTM achieves a 2$\times$ speed-up with only a 0.5% accuracy drop under a 50% token budget, and a 3$\times$ speed-up with just a 2% drop under a 30% budget. Moreover, STTM is query-agnostic, allowing KV cache reuse across different questions for the same video. The project page is available at https://www.jshyun.me/projects/sttm.
title Multi-Granular Spatio-Temporal Token Merging for Training-Free Acceleration of Video LLMs
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2507.07990