Research map / Architectures and efficiency / Mixture of experts
SERE: Similarity-based Expert Re-routing for Efficient Batch Decoding in MoE Models
Juntong Wu, Jialiang Cheng, Fuyu Lv, OU Dan, Li Yuan
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 |
|---|---|---|---|
| re-route to most similar expert instead of a random expert (ablation) | SERE · Re-routing target ablation, top-2 (Qwen1.5-MoE-A2.7B) | average accuracy: 41.68 -> 47.15 | Improved |
| re-route to most similar expert instead of the least similar expert (ablation) | SERE · Re-routing target ablation, top-2 (Qwen1.5-MoE-A2.7B) | average accuracy: 28.74 -> 47.15 | Improved |
| re-route to most similar expert instead of a random expert (ablation) | SERE · Re-routing target ablation, top-1 (Qwen1.5-MoE-A2.7B) | average accuracy: 27.74 -> 37.87 | Improved |
| re-route to most similar expert instead of the least similar expert (ablation) | SERE · Re-routing target ablation, top-1 (Qwen1.5-MoE-A2.7B) | average accuracy: 4.72 -> 37.87 | Improved |
| use Frobenius instead of cosine similarity for the expert similarity matrix (ablation) | SERE · Similarity metric ablation, top-2 (Qwen1.5-MoE-A2.7B) | average accuracy: 45.55 -> 47.15 | Improved |
| use Frobenius instead of CKA-RBF similarity (ablation) | SERE · Similarity metric ablation, top-2 (Qwen1.5-MoE-A2.7B) | average accuracy: 46.85 -> 47.15 | No clear effect |
| use activation-based similarity instead of data-free parameter-based (Concat, Frobenius) similarity (ablation) | SERE · Similarity matrix ablation, top-2 (Qwen1.5-MoE-A2.7B) | average accuracy: 45.14 -> 47.15 | Improved |
| use activation-based similarity instead of data-free parameter-based (Logic, Frobenius) similarity (ablation) | SERE · Similarity matrix ablation, top-2 (Qwen1.5-MoE-A2.7B) | average accuracy: 44.26 -> 47.15 | Improved |
| calibrate with C4 instead of FineWeb-Edu (ablation) | SERE · Calibration dataset ablation, top-2 (Qwen1.5-MoE-A2.7B) | average accuracy: 47.15 -> 47.09 | No clear effect |
| calibrate with domain-specific Exam data instead of FineWeb-Edu at K=1 (ablation) | SERE · Calibration dataset ablation, top-1 (Qwen1.5-MoE-A2.7B) | average accuracy: 37.87 -> 38.14 | Improved |
| increase calibration volume from 200x64 to 800x256 tokens (FineWeb, K=2) (ablation) | SERE · Calibration volume ablation, top-2 (Qwen1.5-MoE-A2.7B) | average accuracy: 46.66 -> 46.58 | No clear effect |
| add critical-expert preservation (rho=0.3) to SERE top-1 (ablation) | SERE · OpenCompass average (Qwen1.5-MoE-A2.7B) | average accuracy: 37.87 -> 39.82 | Improved |
Closest papers by meaning
- $\textit{Read-ME}$: Refactorizing LLMs as Router-Decoupled Mixture of Experts with System Co-Design (NeurIPS 2024 · similarity 0.89)
- Oracle-MoE: Locality-preserving Routing in the Oracle Space for Memory-constrained Large Language Model Inference (ICML 2025 · similarity 0.89)
- FlexMoE: One-for-All Nested Intra-Expert Pruning for MoE Language Models (NeurIPS 2026 · similarity 0.88)
- MoE-SVD: Structured Mixture-of-Experts LLMs Compression via Singular Value Decomposition (ICML 2025 · similarity 0.88)
- MoE++: Accelerating Mixture-of-Experts Methods with Zero-Computation Experts (ICLR 2025 · similarity 0.88)
- LSH-MoE: Communication-efficient MoE Training via Locality-Sensitive Hashing (NeurIPS 2024 · similarity 0.88)
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.
