Research map / Architectures and efficiency / Mixture of experts
Efficient Quantization of Mixture-of-Experts with Theoretical Generalization Guarantees
Mohammed Nowaz Rabbani Chowdhury, Kaoutar El Maghraoui, Hsinyu Tsai, Naigang Wang, Geoffrey W. Burr, Liu Liu, Meng Wang
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 |
|---|---|---|---|
| order experts by router l2-norm change (smaller = higher precision) instead of activation frequency or weights (ablation) | Switch Transformer (CNNDM) mixed precision · CNN/DailyMail | quality retention at low bits: preserved down to 2.5 avg bits vs worse baselines | Improved |
| add MaxVar reordering on top of router-norm ordering (ablation) | Switch Transformer (CNNDM) · CNN/DailyMail | quality retention at low bits: preservation extended to 2.125 bits | Improved |
| use router norm + MaxVar instead of PMQ above 2.0 bits | Mixtral 8x7B · 8 zero-shot tasks avg | average accuracy: 68.38 vs 67.53 at 2.5 bits | Improved |
| use router norm + MaxVar instead of PMQ at 2.0 and 1.75 bits | Mixtral 8x7B · 8 zero-shot tasks avg | average accuracy: 62.56 vs 62.83 at 2.0; 58.95 vs 60.53 at 1.75 | Hurt |
| expert-wise router-norm method vs layer-wise/group-wise methods | Mixtral 8x7B · 8 zero-shot tasks avg | average accuracy: 68.38 vs 49.07 (BSP) at 2.5 bits; 62.56 vs 44.49 (Slim-LLM) at 2.0 | Improved |
| set MaxVar threshold zeta to 3.0 (vs 1.0) (ablation) | Router norm + MaxVar on Mixtral 8x7B · 8 tasks avg, 2.0 bits | average accuracy: 60.44 -> 62.56 | Improved |
| use router norm instead of PMQ for bit assignment | Mixtral 8x7B · assignment overhead | compute time: 2227s and 110 GB GPU -> negligible, no GPU | Improved |
| assign higher precision to less frequently activated experts | Mixtral · Wikitext2 inference time | inference time | Improved |
Closest papers by meaning
- MoEQuant: Enhancing Quantization for Mixture-of-Experts Large Language Models via Expert-Balanced Sampling and Affinity Guidance (ICML 2025 · similarity 0.92)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.91)
- Mixture Compressor for Mixture-of-Experts LLMs Gains More (ICLR 2025 · similarity 0.90)
- Toward Efficient Inference for Mixture of Experts (NeurIPS 2024 · similarity 0.90)
- Pruning and Distilling Mixture-of-Experts into Dense Language Models (NeurIPS 2026 · similarity 0.90)
- Multilinear Mixture of Experts: Scalable Expert Specialization through Factorization (NeurIPS 2024 · 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.
