HybridToken-VLM: Hybrid Token Compression for Vision-Language Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Zhang, Jusheng, Guo, Xiaoyang, Cai, Kaitong, Lv, Qinhan, Fan, Yijia, Chai, Wenhao, Wang, Jian, Wang, Keze
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915662257979392
author Zhang, Jusheng
Guo, Xiaoyang
Cai, Kaitong
Lv, Qinhan
Fan, Yijia
Chai, Wenhao
Wang, Jian
Wang, Keze
author_facet Zhang, Jusheng
Guo, Xiaoyang
Cai, Kaitong
Lv, Qinhan
Fan, Yijia
Chai, Wenhao
Wang, Jian
Wang, Keze
contents Vision-language models (VLMs) have transformed multimodal reasoning, but feeding hundreds of visual patch tokens into LLMs incurs quadratic computational costs, straining memory and context windows. Traditional approaches face a trade-off: continuous compression dilutes high-level semantics such as object identities, while discrete quantization loses fine-grained details such as textures. We introduce HTC-VLM, a hybrid framework that disentangles semantics and appearance through dual channels, i.e., a continuous pathway for fine-grained details via ViT patches and a discrete pathway for symbolic anchors using MGVQ quantization projected to four tokens. These are fused into a 580-token hybrid sequence and compressed into a single voco token via a disentanglement attention mask and bottleneck, ensuring efficient and grounded representations. HTC-VLM achieves an average performance retention of 87.2 percent across seven benchmarks (GQA, VQAv2, MMBench, MME, POPE, SEED-Bench, ScienceQA-Image), outperforming the leading continuous baseline at 81.0 percent with a 580-to-1 compression ratio. Attention analyses show that the compressed token prioritizes the discrete anchor, validating its semantic guidance. Our work demonstrates that a minimalist hybrid design can resolve the efficiency-fidelity dilemma and advance scalable VLMs.
format Preprint
id arxiv_https___arxiv_org_abs_2512_08240
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle HybridToken-VLM: Hybrid Token Compression for Vision-Language Models
Zhang, Jusheng
Guo, Xiaoyang
Cai, Kaitong
Lv, Qinhan
Fan, Yijia
Chai, Wenhao
Wang, Jian
Wang, Keze
Computer Vision and Pattern Recognition
Artificial Intelligence
Vision-language models (VLMs) have transformed multimodal reasoning, but feeding hundreds of visual patch tokens into LLMs incurs quadratic computational costs, straining memory and context windows. Traditional approaches face a trade-off: continuous compression dilutes high-level semantics such as object identities, while discrete quantization loses fine-grained details such as textures. We introduce HTC-VLM, a hybrid framework that disentangles semantics and appearance through dual channels, i.e., a continuous pathway for fine-grained details via ViT patches and a discrete pathway for symbolic anchors using MGVQ quantization projected to four tokens. These are fused into a 580-token hybrid sequence and compressed into a single voco token via a disentanglement attention mask and bottleneck, ensuring efficient and grounded representations. HTC-VLM achieves an average performance retention of 87.2 percent across seven benchmarks (GQA, VQAv2, MMBench, MME, POPE, SEED-Bench, ScienceQA-Image), outperforming the leading continuous baseline at 81.0 percent with a 580-to-1 compression ratio. Attention analyses show that the compressed token prioritizes the discrete anchor, validating its semantic guidance. Our work demonstrates that a minimalist hybrid design can resolve the efficiency-fidelity dilemma and advance scalable VLMs.
title HybridToken-VLM: Hybrid Token Compression for Vision-Language Models
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2512.08240