Harnessing Domain Specialists in Multimodal Mixture-of-Experts for Efficient Adaptation

Paper recorded by Signals 4 on 2026-10-01 in cs.CV. Abstract reproduced from arXiv; link to the original below.

Published 2026-10-01 on arXiv · recorded by Signals 4 on 2026-10-02

Category: cs.CV · 计算机视觉 · first seen 2026-10-02

Abstract

Mixture-of-Experts (MoE) architectures scale model capacity through sparse computation, routing each token through only a small subset of experts. In this work, we explore whether this sparsity gives rise to emergent intrinsic organization in multimodal MoEs. We find that experts develop strong semantic specialization across modalities and domains despite not being explicitly trained for modularit

Read on arXiv →