Research

Paper

AI LLM March 23, 2026

Unified Spatiotemporal Token Compression for Video-LLMs at Ultra-Low Retention

Authors

Junhao Du, Jialong Xue, Anqi Li, Jincheng Dai, Guo Lu

Abstract

Video large language models (Video-LLMs) face high computational costs due to large volumes of visual tokens. Existing token compression methods typically adopt a two-stage spatiotemporal compression strategy, relying on stage-specific metrics and an implicit assumption of spatiotemporal separability. Under extremely low retention ratios, however, such approaches often result in unbalanced allocation and loss of visual evidence essential for question answering. We reformulate token compression as a spatiotemporal allocation task within a global token retention pool. We propose a unified selection mechanism that integrates attention weights and semantic similarity to globally select tokens with high contribution and low redundancy. Unselected tokens are merged via clustering and refilled, preserving information integrity. Inside the LLM, we further introduce text-aware merging to perform secondary compression based on query relevance. Without requiring retraining, our method serves as a plug-and-play module compatible with existing Video-LLMs. Experiments show that retaining only about 2% of visual tokens preserves 90.1% of baseline performance across multiple benchmarks, while reducing FLOPs to roughly 2.6%. These benefits generalize across diverse backbones, decreasing end-to-end inference latency and memory consumption. Our unified spatiotemporal token compression strategy establishes the state-of-the-art in video understanding under ultra-low token retention.

Metadata

arXiv ID: 2603.21957
Provider: ARXIV
Primary Category: cs.CV
Published: 2026-03-23
Fetched: 2026-03-24 06:02

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.21957v1</id>\n    <title>Unified Spatiotemporal Token Compression for Video-LLMs at Ultra-Low Retention</title>\n    <updated>2026-03-23T13:15:22Z</updated>\n    <link href='https://arxiv.org/abs/2603.21957v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.21957v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Video large language models (Video-LLMs) face high computational costs due to large volumes of visual tokens. Existing token compression methods typically adopt a two-stage spatiotemporal compression strategy, relying on stage-specific metrics and an implicit assumption of spatiotemporal separability. Under extremely low retention ratios, however, such approaches often result in unbalanced allocation and loss of visual evidence essential for question answering. We reformulate token compression as a spatiotemporal allocation task within a global token retention pool. We propose a unified selection mechanism that integrates attention weights and semantic similarity to globally select tokens with high contribution and low redundancy. Unselected tokens are merged via clustering and refilled, preserving information integrity. Inside the LLM, we further introduce text-aware merging to perform secondary compression based on query relevance. Without requiring retraining, our method serves as a plug-and-play module compatible with existing Video-LLMs. Experiments show that retaining only about 2% of visual tokens preserves 90.1% of baseline performance across multiple benchmarks, while reducing FLOPs to roughly 2.6%. These benefits generalize across diverse backbones, decreasing end-to-end inference latency and memory consumption. Our unified spatiotemporal token compression strategy establishes the state-of-the-art in video understanding under ultra-low token retention.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CV'/>\n    <published>2026-03-23T13:15:22Z</published>\n    <arxiv:comment>Accepted by CVPR 2026</arxiv:comment>\n    <arxiv:primary_category term='cs.CV'/>\n    <author>\n      <name>Junhao Du</name>\n    </author>\n    <author>\n      <name>Jialong Xue</name>\n    </author>\n    <author>\n      <name>Anqi Li</name>\n    </author>\n    <author>\n      <name>Jincheng Dai</name>\n    </author>\n    <author>\n      <name>Guo Lu</name>\n    </author>\n  </entry>"
}