Research map / Architectures and efficiency / Mixture of experts
SD-MoE: Spectral Decomposition for Effective Expert Specialization
Ruijun Huang, Fang Dong, Xin Zhang, Anrui Chen, Hengjie Cao, Zhendong Huang, Jixian Zhou, Mengyi Chen, Yifeng Yang, Mingzhi Dong, Yujiang Wang, Jinlong Hou and 7 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply SD-MoE spectral decomposition to Qwen MoE | Qwen-2B-A0.8B · 8 downstream tasks avg | average accuracy (%): 50.45 -> 53.97 | Improved |
| apply SD-MoE to DeepSeek MoE | DeepSeek-2B-A0.8B · 8 downstream tasks avg | average accuracy (%): 50.51 -> 51.49 | Improved |
| apply SD-MoE to larger Qwen MoE | Qwen-7B-A1.5B · 8 downstream tasks avg | average accuracy (%): 55.34 -> 57.23 | Improved |
| apply SD-MoE (LAMBADA exception at large scale) | Qwen-7B-A1.5B · LAMBADA_OPENAI | accuracy (%): 49.31 -> 47.29 | Hurt |
| add gradient projection to parameter decomposition (ablation) | SD-MoE · 8 downstream tasks avg | average accuracy (%): 48.54 -> 48.86 (baseline); full method needed for gain | Improved |
| vary common subspace rank k in {2,4,8,16,32} (ablation) | SD-MoE · 8 downstream tasks avg | average accuracy (%): 53.64, 53.39, 53.97, 53.57, 53.68 | No clear effect |
| use SD-MoE instead of AES diversity regularization | Qwen-2B-A0.8B · 8 downstream tasks avg | average accuracy (%): 51.58 -> 53.97 | Improved |
| apply SD-MoE for summarization | Qwen MoE · XSUM | ROUGE-1: 33.10 -> 34.27 | Improved |
| apply SD-MoE to allow larger learning rate | Qwen MoE · learning rate stability | max stable learning rate: 1e-4 -> 4e-4 | Improved |
| apply SD-MoE (training cost) | Qwen-2B-A0.8B · training throughput | tokens/sec: 229K -> 218K | Hurt |
| apply SD-MoE to reduce inter-expert subspace similarity | Qwen MoE · expert principal subspace similarity | similarity: ~0.7 avg -> ~0.1 | Improved |
| periodic SVD refresh interval 8 vs 16 vs 32 steps (ablation) | SD-MoE · training overhead | SVD refresh overhead: 0.52% / 0.46% / 0.35% | No clear effect |
Closest papers by meaning
- Advancing Expert Specialization for Better MoE (NeurIPS 2025 · similarity 0.92)
- Pruning and Distilling Mixture-of-Experts into Dense Language Models (NeurIPS 2026 · similarity 0.90)
- FourierMoE: Fourier Mixture-of-Experts Adaptation of Large Language Models (NeurIPS 2026 · similarity 0.89)
- BAM! Just Like That: Simple and Efficient Parameter Upcycling for Mixture of Experts (NeurIPS 2024 · similarity 0.89)
- On the Expressive Power of Mixture-of-Experts for Structured Complex Tasks (NeurIPS 2025 · similarity 0.89)
- Mixture-of-Experts Meets Instruction Tuning: A Winning Combination for Large Language Models (ICLR 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
