Research map / Architectures and efficiency / Mixture of experts
Attribution-Guided and Coverage-Maximized Pruning for Structural MoE Compression
Yifu Ding, jiacheng wang, Ge Yang, Yongcheng Jing, Jinyang Guo, Xianglong Liu, Dacheng Tao
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 |
|---|---|---|---|
| use attribution-based intra-layer allocation instead of uniform (ablation) | coverage-based pruning · GSM8K (Qwen1.5-MoE-A2.7B, 50% prune) | accuracy: 52.6 -> 58.2 | Improved |
| use attribution instead of raw loss in intra-layer coverage (ablation) | coverage-based pruning · GSM8K (Qwen1.5-MoE-A2.7B, 50% prune) | accuracy: 45.7 -> 58.2 | Improved |
| use attribution instead of raw loss in intra-layer coverage (ablation) | coverage-based pruning · HumanEval (Qwen1.5-MoE-A2.7B, 50% prune) | pass@1: 12.2 -> 30.5 | Improved |
| use attribution instead of router logits (ablation) | coverage-based pruning · GSM8K (Qwen1.5-MoE-A2.7B, 50% prune) | accuracy: 55.2 -> 58.2 | Improved |
| use coverage-based instead of top-k channel allocation (ablation) | intra-layer allocation · GSM8K (Qwen1.5-MoE-A2.7B, 50% prune) | accuracy: 50.6 -> 58.2 | Improved |
| use smoothed loss instead of uniform for inter-layer coverage (ablation) | coverage-based pruning · GSM8K (Qwen1.5-MoE-A2.7B, 50% prune) | accuracy: 57.0 -> 58.2 | Improved |
| use smoothed loss instead of raw loss for inter-layer coverage (ablation) | coverage-based pruning · GSM8K (Qwen1.5-MoE-A2.7B, 50% prune) | accuracy: 56.1 -> 58.2 | Improved |
| use smoothed loss instead of uniform (ablation) | coverage-based pruning · HumanEval (Qwen1.5-MoE-A2.7B, 50% prune) | pass@1: 29.3 -> 30.5 | No clear effect |
| add square-root smoothing of layerwise loss (ablation) | coverage allocation · Avg (Qwen1.5-MoE-A2.7B, 50% prune) | average accuracy: 52.40 -> 58.27 | Improved |
| sqrt instead of log smoothing (ablation) | coverage allocation · Avg (Qwen1.5-MoE-A2.7B, 50% prune) | average accuracy: 56.92 -> 58.27 | Improved |
| largest removed scores vs largest removed channels for AAR (ablation) | alignment-aware redistribution · CSQA Avg (Qwen1.5-MoE-A2.7B) | accuracy: 65.47 -> 66.40 | No clear effect |
| switch routing from top-2 to top-1 experts | pruned Qwen1.5-MoE-A2.7B (50%) · Avg | average accuracy: 58.27 -> 55.15 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- PuzzleMoE 50% — PuzzleMoE: Efficient Compression of Large Mixture-of-Experts Models via Sparse Expert Merging and Bit-packed inference (ICML 2026) · better on 3 of 3 shared results
- C-Prune 25% + 4b — CPrune: Compiler-Informed Model Pruning for Efficient Target-Aware DNN Execution (ECCV 2022) · better on 2 of 2 shared results
- MoNE 50% — MoNE: Replacing Redundant Experts with Lightweight Novices for Structured Pruning of MoE (ICLR 2026) · better on 1 of 1 shared result
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
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.91)
- Pruning and Distilling Mixture-of-Experts into Dense Language Models (NeurIPS 2026 · similarity 0.90)
- FlexMoE: One-for-All Nested Intra-Expert Pruning for MoE Language Models (NeurIPS 2026 · similarity 0.90)
- DiEP: Adaptive Mixture-of-Experts Compression through Differentiable Expert Pruning (NeurIPS 2025 · similarity 0.90)
- Domain-Specific Pruning of Large Mixture-of-Experts Models with Few-shot Demonstrations (NeurIPS 2025 · similarity 0.90)
- Mixture Compressor for Mixture-of-Experts LLMs Gains More (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.
