SD-MoE: Spectral Decomposition for Effective Expert Specialization
This work addresses a bottleneck in scaling large language models via MoE architectures, offering an incremental improvement for enhancing expert specialization.
The paper tackled the problem of ineffective expert specialization in Mixture-of-Experts architectures, where experts become functionally similar, limiting model performance, and proposed Spectral-Decoupled MoE (SD-MoE) to address this, resulting in improved performance across downstream tasks with minimal additional computation.
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.