Research map / Architectures and efficiency / Mixture of experts
Less Token, More Signal: MoE Expert Pruning via Critical Token Selection
Zeliang Zong, Kai Zhang, Yarong Wang, Wenming Tan, Ye Ren, Jilin Hu
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 |
|---|---|---|---|
| add attention-guided token selection to frequency-based pruning (ablation) | frequency-based expert pruning on OLMoE-7A1B · 8 zero-shot tasks avg (25% pruned) | accuracy: 53.48 -> 55.12 | Improved |
| add dual-factor scoring (ablation) | OLMoE-7A1B pruning · 8 zero-shot tasks avg (25% pruned) | accuracy: 53.48 -> 55.43 | Improved |
| add experts-to-bias conversion on top of dual-factor scoring (ablation) | OLMoE-7A1B pruning · 8 zero-shot tasks avg (25% pruned) | accuracy: 55.43 -> 56.16 | Improved |
| add token selection on top of dual-factor scoring and experts-to-bias (ablation) | OLMoE-7A1B pruning · 8 zero-shot tasks avg (25% pruned) | accuracy: 56.16 -> 58.42 (+2.26) | Improved |
| add experts-to-bias stage (Stage 3) to STEP (ablation) | STEP on OLMoE-7A1B · 8 zero-shot tasks avg (25% pruned) | accuracy: 57.51 -> 58.42 | Improved |
| add experts-to-bias stage (Stage 3) to STEP (ablation) | STEP on OLMoE-7A1B · 8 zero-shot tasks avg (50% pruned) | accuracy: 48.31 -> 50.05 | Improved |
| add experts-to-bias stage (Stage 3) to STEP (ablation) | STEP on Qwen3-30A3B · 8 zero-shot tasks avg (25% pruned) | accuracy: 68.84 -> 69.06 | No clear effect |
| reduce token retention from 100% to 50% (ablation) | STEP on OLMoE-7A1B · WikiText perplexity (25% pruned) | perplexity: 11.65 -> 10.40 | Improved |
| reduce token retention from 50% to 20% (ablation) | STEP on OLMoE-7A1B · 8 zero-shot tasks avg (25% pruned) | accuracy: 58.42 -> 58.81 | No clear effect |
| use gradient-based token importance instead of attention (ablation) | STEP token selection · 8 zero-shot tasks avg (OLMoE-7A1B) | accuracy gain: +0.42 and +0.24 vs +2.26 and +2.65 | No clear effect |
| use L2 norm or update magnitude token scoring instead of attention (ablation) | STEP token selection · 8 zero-shot tasks avg (OLMoE-7A1B) | accuracy gain: -3.06, -4.54, -2.72, -3.09 | Hurt |
| use LLMLingua token selection instead of attention (ablation) | STEP token selection · 8 zero-shot tasks avg (OLMoE-7A1B) | accuracy: 58.81 -> 57.22 | Hurt |
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.
- MoNE — MoNE: Replacing Redundant Experts with Lightweight Novices for Structured Pruning of MoE (ICLR 2026) · better on 9 of 9 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
- Pruning and Distilling Mixture-of-Experts into Dense Language Models (NeurIPS 2026 · similarity 0.92)
- Domain-Specific Pruning of Large Mixture-of-Experts Models with Few-shot Demonstrations (NeurIPS 2025 · similarity 0.92)
- Advancing Expert Specialization for Better MoE (NeurIPS 2025 · similarity 0.92)
- 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)
- Solving Token Gradient Conflict in Mixture-of-Experts for Large Vision-Language Model (ICLR 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.
