SMoES: Soft Modality-Guided Expert Specialization in MoE-VLMs

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Bo, Zi-Hao, Li, Yaqian, Hou, Anzhou, Takezoe, Rinyoichi, Zhao, Ertao, Pan, Tianxiang, Yan, Jiale, Guang, Mo, Long, Kaiwen
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918469414420480
author Bo, Zi-Hao
Li, Yaqian
Hou, Anzhou
Takezoe, Rinyoichi
Zhao, Ertao
Pan, Tianxiang
Yan, Jiale
Guang, Mo
Long, Kaiwen
author_facet Bo, Zi-Hao
Li, Yaqian
Hou, Anzhou
Takezoe, Rinyoichi
Zhao, Ertao
Pan, Tianxiang
Yan, Jiale
Guang, Mo
Long, Kaiwen
contents Mixture-of-Experts (MoE) has become a prevalent backbone for large vision-language models (VLMs), yet how modality-specific signals should guide expert routing remains under-explored. Existing routing strategies are either hand-crafted or modality-agnostic, relying on idealized priors that ignore the layer-dependent modality fusion patterns in MoE-VLMs and provide little guidance for expert specialization. We propose Soft Modality-guided Expert Specialization (SMoES), which consists of dynamic soft modality scores that capture layer-dependent fusion patterns, an expert binning mechanism aligned with expert-parallel deployment, and an inter-bin mutual information regularization that encourages coherent modality specialization. Our method leverages attention-based or Gaussian-statistics modality scores to optimize mutual information regularization. Experiments across four MoE-based VLMs and 16 benchmarks demonstrate improvement on both effectiveness and efficiency: 0.9% and 4.2% average gain on multimodal and language-only tasks, 56.1% reduction in EP communication overhead, and 12.3% throughput improvement under realistic deployment. These results validate that aligning routing with modality-aware expert specialization unlocks MoE-VLM capacity and efficiency.
format Preprint
id arxiv_https___arxiv_org_abs_2604_23996
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle SMoES: Soft Modality-Guided Expert Specialization in MoE-VLMs
Bo, Zi-Hao
Li, Yaqian
Hou, Anzhou
Takezoe, Rinyoichi
Zhao, Ertao
Pan, Tianxiang
Yan, Jiale
Guang, Mo
Long, Kaiwen
Computer Vision and Pattern Recognition
Mixture-of-Experts (MoE) has become a prevalent backbone for large vision-language models (VLMs), yet how modality-specific signals should guide expert routing remains under-explored. Existing routing strategies are either hand-crafted or modality-agnostic, relying on idealized priors that ignore the layer-dependent modality fusion patterns in MoE-VLMs and provide little guidance for expert specialization. We propose Soft Modality-guided Expert Specialization (SMoES), which consists of dynamic soft modality scores that capture layer-dependent fusion patterns, an expert binning mechanism aligned with expert-parallel deployment, and an inter-bin mutual information regularization that encourages coherent modality specialization. Our method leverages attention-based or Gaussian-statistics modality scores to optimize mutual information regularization. Experiments across four MoE-based VLMs and 16 benchmarks demonstrate improvement on both effectiveness and efficiency: 0.9% and 4.2% average gain on multimodal and language-only tasks, 56.1% reduction in EP communication overhead, and 12.3% throughput improvement under realistic deployment. These results validate that aligning routing with modality-aware expert specialization unlocks MoE-VLM capacity and efficiency.
title SMoES: Soft Modality-Guided Expert Specialization in MoE-VLMs
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2604.23996