Research map / Architectures and efficiency / Mixture of experts
C-GNN-PRUNE: A Unified Graph-Based Framework for Structure-Aware Pruning of Mixture-of-Experts Models
Lin Li, Yan Wang, Zhuopeng 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 |
|---|---|---|---|
| add GNN structure-aware embedding module (ablation) | C-GNN-PRUNE · MMLU Computer Science (Qwen1.5-MoE-A2.7B, 20% pruning) | accuracy: 44.99 -> 50.50 | Improved |
| add GNN structure-aware embedding module (ablation) | C-GNN-PRUNE · MMLU Business (Qwen1.5-MoE-A2.7B, 20% pruning) | accuracy: 42.18 -> 45.69 | Improved |
| add GNN structure-aware embedding module (ablation) | C-GNN-PRUNE · MMLU average (Qwen1.5-MoE-A2.7B) | accuracy: 40.59 -> 43.96 | Improved |
| add entropy-guided layer-wise pruning allocation instead of uniform ratio (ablation) | C-GNN-PRUNE · MMLU Computer Science (Qwen1.5-MoE-A2.7B, 20% pruning) | accuracy: 45.49 -> 50.50 | Improved |
| add entropy-guided layer-wise pruning allocation (ablation) | C-GNN-PRUNE · MMLU Math (Qwen1.5-MoE-A2.7B, 20% pruning) | accuracy: 34.68 -> 35.68 | No clear effect |
| mix output similarity and routing divergence in graph (omega 0.3) instead of output-only (omega 0.0) (ablation) | C-GNN-PRUNE · ARC-e (Mixtral 8x7B, 20% pruning) | accuracy: 78.87 -> 80.47 | Improved |
| mix output similarity and routing divergence (omega 0.3) instead of routing-only (omega 1.0) (ablation) | C-GNN-PRUNE · BoolQ (Mixtral 8x7B, 20% pruning) | accuracy: 76.26 -> 84.49 | Improved |
| use Girvan-Newman community detection instead of KMeans (ablation) | C-GNN-PRUNE · BoolQ (Mixtral 8x7B) | accuracy: 80.52 -> 84.49 | Improved |
| use Girvan-Newman instead of Agglomerative clustering (ablation) | C-GNN-PRUNE · BoolQ and WinoGrande (Mixtral 8x7B) | accuracy: BoolQ 85.50 vs 84.49; ARC-e 78.87 vs 80.47 | Mixed |
| C-GNN-PRUNE vs C-PRUNE at 20% pruning | C-GNN-PRUNE · DeepSeek-V2-Lite MMLU Computer Science | accuracy: 51.50 -> 50.10 | Hurt |
| C-GNN-PRUNE vs C-PRUNE at 20% pruning | C-GNN-PRUNE · Qwen1.5-MoE-A2.7B average | score: 37.65 -> 43.38 | Improved |
| C-GNN-PRUNE vs EEMoE at 50% pruning | C-GNN-PRUNE · Mixtral 8x7B average over seven benchmarks | accuracy: 61.10 -> 62.49 | 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.
- C-PRUNE — CPrune: Compiler-Informed Model Pruning for Efficient Target-Aware DNN Execution (ECCV 2022) · better on 6 of 6 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
- FlexMoE: One-for-All Nested Intra-Expert Pruning for MoE Language Models (NeurIPS 2026 · similarity 0.91)
- Pruning and Distilling Mixture-of-Experts into Dense Language Models (NeurIPS 2026 · similarity 0.91)
- 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)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.90)
- Advancing Expert Specialization for Better MoE (NeurIPS 2025 · similarity 0.89)
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.
