AI Paper Atlas

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

ICML 2026 · Architectures and efficiency › Mixture of experts · approach: MoE LLM compression · 0 citations (OpenAlex)

Read the paper

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 changedWhereMeasuredEffect
use RFID-MoE rank allocation and residual reconstruction instead of MoBEMoE SVD compression · PTB, Qwen3-30B-A3B-2507, 60% compressionperplexity: 24.93 -> 16.92Improved
use RFID-MoE instead of MoBEMoE SVD compression · PTB, DeepSeekMoE-16B-Base, 60% compressionperplexity: 11.90 -> 11.54Improved
use RFID-MoE instead of MoBEMoE SVD compression · PTB, Qwen3-235B-A22B-2507perplexity: 11.13 -> 10.67Improved
use RFID-MoE instead of MoBE at 20% compressionMoE SVD compression · PTB, Qwen3-30B-A3B-2507, 20% compressionperplexity: 12.73 -> 12.53Improved
use RFID-MoE instead of MoBEMoE SVD compression · PTB, Ling-mini-2.0, 40% compressionperplexity: 25.48 -> 24.18Improved
use RFID-MoE instead of MoBEMoE SVD compression · Zero-shot average accuracy, Qwen2-57B-A14B, 40% compressionaverage accuracy: 0.66 -> 0.67No clear effect
use RFID-MoE instead of MoBEMoE SVD compression · Zero-shot average accuracy, Qwen3-30B-A3B-2507, 60% compressionaverage accuracy: 0.61 -> 0.64Improved
lower fusion weight xi from 0.9 to 0.7 (more routing frequency) (ablation)RFID-MoE · PTB, 40% compressionperplexity: 13.09 -> 12.81Improved
raise fusion weight xi from 0.5 to 0.7 (ablation)RFID-MoE · PTB, 40% compressionperplexity: 12.86 -> 12.81Improved
vary xi at 20% compression (0.9, 0.8, 0.7, 0.5) (ablation)RFID-MoE · PTB, 20% compressionperplexity: 12.58 -> 12.53No clear effect
calibrate routing frequency on C4 instead of WikiText-2 (ablation)RFID-MoE · Qwen3-30B-A3B-2507, 60% compressionperplexity: WikiText-2 9.57 -> 10.51; PTB 16.92 -> 13.70; C4 17.96 -> 16.85Mixed
use RFID-MoE instead of MoBE on a smaller modelMoE SVD compression · PTB, Qwen1.5-MoE-A2.7B, 60% compressionperplexity: 13.82 -> 13.66Improved

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.

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

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy