Research map / Learning theory and algorithms / Prediction and decision losses
Geometric and Stochastic Analysis of Discontinuities in Sparse Mixture-of-Experts
Tho Tran Huu, Huu-Tuan Nguyen, Thien-Hai Nguyen, Nhat-Tri Ho, Viet-Hoang Tran, Tho Quan, Tan Minh Nguyen
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 l-inf epsilon local smoothing of gating logits near top-k boundaries | Switch Transformer SMoE · WikiText-103 test | perplexity: 35.52 -> 34.35 | Improved |
| add l-inf epsilon local smoothing | Switch Transformer SMoE · Attacked WikiText-103 test | perplexity: 44.18 -> 42.85 | Improved |
| add smoothing instead of ReMoE continuous routing | Switch Transformer SMoE · WikiText-103 test | perplexity: 35.35 -> 34.35 | Improved |
| add smoothing | Switch Transformer SMoE · EnWiki-8 | bits per character: 1.153 -> 1.122 | Improved |
| add smoothing (k=2) | BERT-large with one MoE layer · GLUE average (5 tasks) | average score: 81.17 -> 81.65 | Improved |
| add smoothing (k=4) | BERT-large with one MoE layer · GLUE average (5 tasks) | average score: 81.14 -> 81.73 | Improved |
| add smoothing (k=2) | BERT-large with one MoE layer · GLUE QNLI | accuracy: 92.56 -> 92.40 | No clear effect |
| add smoothing to GMoE | GMoE ViT-S/16 · DomainBed average | accuracy (%): 66.84 -> 67.4 | Improved |
| add smoothing to GMoE | GMoE ViT-S/16 · DomainBed PACS | accuracy (%): 87.7 -> 87.6 | No clear effect |
| anneal smoothing away during training and use hard top-2 at inference (ablation) | SmoothSMoE · WikiText-103 test | perplexity: 35.52 -> 34.62 | Improved |
| keep smoothing at inference (k=2.5) versus annealed hard top-2 (ablation) | SmoothSMoE · WikiText-103 test | perplexity: 34.62 -> 34.35 | Improved |
| add smoothing (k=2.5) to upcycled VLM | 5.6B ViT-based VLM, 4 experts · Vision-language average over 10 benchmarks | average score: 54.53 -> 55.07 | Improved |
Closest papers by meaning
- Merge, Then Compress: Demystify Efficient SMoE with Hints from Its Routing Policy (ICLR 2024 · similarity 0.87)
- Routers Learn the Geometry of Their Experts: Geometric Coupling in Sparse Mixture-of-Experts (NeurIPS 2026 · similarity 0.87)
- On the Expressive Power of Mixture-of-Experts for Structured Complex Tasks (NeurIPS 2025 · similarity 0.86)
- A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts (ICML 2024 · similarity 0.86)
- Tight Clusters Make Specialized Experts (ICLR 2025 · similarity 0.86)
- Retraining-free Merging of Sparse MoE via Hierarchical Clustering (ICML 2025 · similarity 0.86)
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.
