Research map / Architectures and efficiency / Mixture of experts
Effective MoE-based LLM Compression by Exploiting Heterogeneous Inter-Group Experts Routing Frequency and Information Density
Zhendong Mi, Yixiao Chen, Pu Zhao, Xiaodong Yu, Hao Wang, Yanzhi Wang, Shaoyi Huang
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 |
|---|---|---|---|
| use RFID-MoE rank allocation and residual reconstruction instead of MoBE | MoE SVD compression · PTB, Qwen3-30B-A3B-2507, 60% compression | perplexity: 24.93 -> 16.92 | Improved |
| use RFID-MoE instead of MoBE | MoE SVD compression · PTB, DeepSeekMoE-16B-Base, 60% compression | perplexity: 11.90 -> 11.54 | Improved |
| use RFID-MoE instead of MoBE | MoE SVD compression · PTB, Qwen3-235B-A22B-2507 | perplexity: 11.13 -> 10.67 | Improved |
| use RFID-MoE instead of MoBE at 20% compression | MoE SVD compression · PTB, Qwen3-30B-A3B-2507, 20% compression | perplexity: 12.73 -> 12.53 | Improved |
| use RFID-MoE instead of MoBE | MoE SVD compression · PTB, Ling-mini-2.0, 40% compression | perplexity: 25.48 -> 24.18 | Improved |
| use RFID-MoE instead of MoBE | MoE SVD compression · Zero-shot average accuracy, Qwen2-57B-A14B, 40% compression | average accuracy: 0.66 -> 0.67 | No clear effect |
| use RFID-MoE instead of MoBE | MoE SVD compression · Zero-shot average accuracy, Qwen3-30B-A3B-2507, 60% compression | average accuracy: 0.61 -> 0.64 | Improved |
| lower fusion weight xi from 0.9 to 0.7 (more routing frequency) (ablation) | RFID-MoE · PTB, 40% compression | perplexity: 13.09 -> 12.81 | Improved |
| raise fusion weight xi from 0.5 to 0.7 (ablation) | RFID-MoE · PTB, 40% compression | perplexity: 12.86 -> 12.81 | Improved |
| vary xi at 20% compression (0.9, 0.8, 0.7, 0.5) (ablation) | RFID-MoE · PTB, 20% compression | perplexity: 12.58 -> 12.53 | No clear effect |
| calibrate routing frequency on C4 instead of WikiText-2 (ablation) | RFID-MoE · Qwen3-30B-A3B-2507, 60% compression | perplexity: WikiText-2 9.57 -> 10.51; PTB 16.92 -> 13.70; C4 17.96 -> 16.85 | Mixed |
| use RFID-MoE instead of MoBE on a smaller model | MoE SVD compression · PTB, Qwen1.5-MoE-A2.7B, 60% compression | perplexity: 13.82 -> 13.66 | Improved |
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.
- MoBE — MoBE: Mixture-of-Basis-Experts for Compressing MoE-based LLMs (ICLR 2026) · better on 10 of 10 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
- Mixture Compressor for Mixture-of-Experts LLMs Gains More (ICLR 2025 · similarity 0.93)
- MoE-SVD: Structured Mixture-of-Experts LLMs Compression via Singular Value Decomposition (ICML 2025 · similarity 0.92)
- Delta Decompression for MoE-based LLMs Compression (ICML 2025 · similarity 0.91)
- FlexMoE: One-for-All Nested Intra-Expert Pruning for MoE Language Models (NeurIPS 2026 · similarity 0.91)
- LightMoE: Reducing Mixture-of-Experts Redundancy through Expert Replacing (NeurIPS 2026 · similarity 0.90)
- Pruning and Distilling Mixture-of-Experts into Dense Language Models (NeurIPS 2026 · 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.
