Research map / Architectures and efficiency / Mixture of experts
TD-MoE: Tensor Decomposition for MoE Models
Yuebin XU, YANHONG WANG, Xuemei Peng, Hui Zang, Chen Minghao, Pengfei Xia, Zeyi Wen
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 |
|---|---|---|---|
| replace per-expert SVD with joint Tucker decomposition (output whitening) | MoE-SVD, Qwen2-57B-A14B · 40% compression avg accuracy | avg accuracy: 0.48 -> 0.56 | Improved |
| replace per-expert SVD with joint Tucker decomposition (output whitening) | MoE-SVD, Mixtral-8x7B · 40% compression avg accuracy | avg accuracy: 0.50 -> 0.57 | Improved |
| replace per-expert SVD with joint Tucker decomposition (input whitening) | MoE-SVD, Mixtral-8x7B · 60% compression avg accuracy | avg accuracy: 0.37 -> 0.45 | Improved |
| replace per-expert SVD with joint Tucker decomposition (input whitening) | MoE-SVD, Qwen2-57B-A14B · 60% compression avg accuracy | avg accuracy: 0.47 -> 0.52 | Improved |
| replace per-expert SVD with joint Tucker decomposition (output whitening) | MoE-SVD, Mixtral-8x7B · 20% compression WikiText-2 | perplexity: 4.86 -> 4.49 | Improved |
| use input whitening instead of output whitening (ablation) | TD-MoE, Mixtral-8x7B · 60% compression avg accuracy | avg accuracy: 0.42 -> 0.45 | Improved |
| use output whitening instead of input whitening (ablation) | TD-MoE, Mixtral-8x7B · 40% compression avg accuracy | avg accuracy: 0.54 -> 0.57 | Improved |
| add NF4 quantization after decomposition | TD-MoE, Mixtral-8x7B · 20% compression | LM avg perplexity: 9.81 -> 9.70 | No clear effect |
| add 20% structured pruning of Tucker core | TD-MoE, Mixtral-8x7B · 20% compression | reasoning avg accuracy: 0.66 -> 0.64 | Hurt |
| add 40% structured pruning of Tucker core | TD-MoE, Mixtral-8x7B · 20% compression | reasoning avg accuracy: 0.66 -> 0.61 | Hurt |
| add multi-linear whitening | TD-MoE · covariance conditioning | post-whitening deviation: 9 - 5.4x10^4 -> 1.1x10^-7 | Improved |
| vary calibration set size N from 128 to 2k (ablation) | TD-MoE · 20% compression | avg PPL: 9.81-9.84 | No clear effect |
Later papers that compared with this method
1 later accepted paper reports TD-MoE as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- FlexMoE: One-for-All Nested Intra-Expert Pruning for MoE Language Models (NeurIPS 2026 · better on 9 of 9 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MoE-SVD — MoE-SVD: Structured Mixture-of-Experts LLMs Compression via Singular Value Decomposition (ICML 2025) · better on 13 of 13 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- MoE-SVD: Structured Mixture-of-Experts LLMs Compression via Singular Value Decomposition (ICML 2025 · similarity 0.92)
- Pruning and Distilling Mixture-of-Experts into Dense Language Models (NeurIPS 2026 · similarity 0.91)
- FlexMoE: One-for-All Nested Intra-Expert Pruning for MoE Language Models (NeurIPS 2026 · similarity 0.91)
- Mixture Compressor for Mixture-of-Experts LLMs Gains More (ICLR 2025 · similarity 0.90)
- LightMoE: Reducing Mixture-of-Experts Redundancy through Expert Replacing (NeurIPS 2026 · similarity 0.90)
- DeRS: Towards Extremely Efficient Upcycled Mixture-of-Experts Models (CVPR 2025 · similarity 0.90)
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.
