SD-MoE: Spectral Decomposition for Effective Expert Specialization

TL;DR

SD-MoE introduces a spectral decomposition approach to enhance expert specialization in Mixture-of-Experts models, addressing spectral overlap issues to improve model capacity and performance with minimal extra computation.

Contribution

The paper presents a novel spectral decomposition method for MoE models that improves expert specialization and overall performance by addressing spectral overlap and alignment issues.

Findings

Spectral overlap among experts limits specialization.

Spectral decomposition improves expert differentiation.

SD-MoE enhances downstream task performance.

Abstract

Mixture-of-Experts (MoE) architectures scale Large Language Models via expert specialization induced by conditional computation. In practice, however, expert specialization often fails: some experts become functionally similar, while others functioning as de facto shared experts, limiting the effective capacity and model performance. In this work, we analysis from a spectral perspective on parameter and gradient spaces, uncover that (1) experts share highly overlapping dominant spectral components in their parameters, (2) dominant gradient subspaces are strongly aligned across experts, driven by ubiquitous low-rank structure in human corpus, and (3) gating mechanisms preferentially route inputs along these dominant directions, further limiting specialization. To address this, we propose Spectral-Decoupled MoE (SD-MoE), which decomposes both parameter and gradient in the spectral space. SD-MoE improves performance across downstream tasks, enables effective expert specialization, incurring minimal additional computation, and can be seamlessly integrated into a wide range of existing MoE architectures, including Qwen and DeepSeek.