Research map / Architectures and efficiency / Mixture of experts
SCHUR-A*: Layer-wise Optimal Expert Pruning for MoEs via Schur-Complement Guided A* Search
Zheng Chen, Yang Weifeng, Jianxiao Tang, Buhui Yao
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 original gate weights (Mask-Only) instead of refitted coefficients at inference (ablation) | SCHUR-A* on Qwen3-30B-A3B · PPL | perplexity: 8.84 -> 8.25 | Improved |
| use original gate weights (Mask-Only) instead of refitted coefficients at inference (ablation) | SCHUR-A* on ERNIE-4.5-21B-A3B · PPL | perplexity: 8.12 -> 7.55 | Improved |
| use original gate weights (Mask-Only) instead of refitted coefficients at inference (ablation) | SCHUR-A* on Qwen3-30B-A3B · MMLU | accuracy: 0.579 -> 0.623 | Improved |
| use original gate weights (Mask-Only) instead of refitted coefficients at inference (ablation) | SCHUR-A* on Qwen3-30B-A3B · 8-task average | average accuracy: 0.658 -> 0.661 | Improved |
| add A* global search instead of greedy selection with same gain metric (ablation) | SCHUR-A* on Qwen3-30B-A3B · PPL | perplexity: 10.30 -> 8.25 | Improved |
| add A* global search instead of greedy selection with same gain metric (ablation) | SCHUR-A* on Qwen3-30B-A3B · MMLU | accuracy: 0.580 -> 0.623 | Improved |
| add A* global search instead of greedy selection with same gain metric (ablation) | SCHUR-A* on Qwen3-30B-A3B · ARC-c | accuracy: 0.449 -> 0.491 | Improved |
| add A* global search instead of greedy selection with same gain metric (ablation) | SCHUR-A* on Qwen3-30B-A3B · OpenBookQA | accuracy: 0.432 -> 0.452 | Improved |
| add adaptive Kneedle layer-wise budgeting instead of fixed 50% pruning (ablation) | SCHUR-A* on Qwen3-30B-A3B-Instruct-2507 · MMLU | accuracy: 0.550 -> 0.672 | Improved |
| add adaptive Kneedle layer-wise budgeting instead of fixed 50% pruning (ablation) | SCHUR-A* on Qwen3-30B-A3B-Instruct-2507 · GSM8K | accuracy: +19.6% | Improved |
| use SCHUR-A* criterion instead of REAP at the same fixed 42% pruning ratio (ablation) | expert pruning on Qwen3-30B-A3B · 8-task average | average accuracy: 0.645 -> 0.661 | Improved |
| apply Kneedle adaptive budget to REAP baseline (ablation) | REAP on Qwen3-30B-A3B · 8-task average | average accuracy: 0.645 -> 0.642 | No clear effect |
Closest papers by meaning
- Towards Principled Fine-Grained MoE Expert Pruning via Pseudo-Boolean Approximation (NeurIPS 2026 · similarity 0.93)
- Pruning and Distilling Mixture-of-Experts into Dense Language Models (NeurIPS 2026 · similarity 0.90)
- DiEP: Adaptive Mixture-of-Experts Compression through Differentiable Expert Pruning (NeurIPS 2025 · similarity 0.90)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.90)
- Domain-Specific Pruning of Large Mixture-of-Experts Models with Few-shot Demonstrations (NeurIPS 2025 · similarity 0.90)
- FlexMoE: One-for-All Nested Intra-Expert Pruning for MoE Language Models (NeurIPS 2026 · 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.
