Research map / Architectures and efficiency / Mixture of experts
SoftMoE: Soft Differentiable Routing for Mixture-of-Experts in LLMs
Mikołaj Zasada, Łukasz Struski, Jacek Tabor, Marcin Kurdziel
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 soft top-k routing (SoftMoE*) instead of sparse top-2 routing | Sparse MoE k=2 · OWT validation | loss: 2.79 -> 2.78 | Improved |
| use soft top-k routing (SoftMoE*) instead of sparse top-2 routing | Sparse MoE k=2 · C4 validation | loss: 2.24 -> 2.23 | Improved |
| use SoftMoE* k=1 alpha=4 instead of sparse top-4 | Sparse MoE k=4 · OWT validation | loss: 2.75 -> 2.70 | Improved |
| use SoftMoE* k=1 alpha=4 instead of sparse top-4 | Sparse MoE k=4 · C4 HellaSwag | accuracy: 47.23 -> 48.48 | Improved |
| add learned expert allocation across layers (ablation) | SoftMoE* k=1.5 alpha=2 · OWT validation | loss: 2.78 -> 2.76 | Improved |
| add learned expert allocation across layers (ablation) | SoftMoE* k=2 alpha=2 · OWT validation | loss: 2.75 -> 2.74 | Improved |
| add learned expert allocation across layers (ablation) | SoftMoE* k=2 alpha=2 · C4 validation | loss: 2.21 -> 2.20 | Improved |
| add learned expert allocation across layers (ablation) | SoftMoE* k=2 alpha=2 · C4 HellaSwag | accuracy: 48.05 -> 47.48 | Hurt |
| lower inference truncation threshold tau from 2.4 to 1.2 (ablation) | SoftMoE k=2 · C4 HellaSwag | accuracy: 46.98 -> 47.48 | Mixed |
| use truncation tau 1.3 instead of 1.5 (ablation) | SoftMoE* k=1.5 alpha=2 · OWT PIQA | accuracy: 62.02 -> 63.87 | Improved |
| add routing jitter 0.05 to sparse MoE top-2 (ablation) | Sparse MoE k=2 · OWT-style small run | loss: 2.99 -> 3.00 | Hurt |
| add routing jitter 0.05 to sparse MoE top-1 (ablation) | Sparse MoE k=1 · HellaSwag | accuracy: 28.79 -> 29.21 | No clear effect |
Closest papers by meaning
- Routing Experts: Learning to Route Dynamic Experts in Existing Multi-modal Large Language Models (ICLR 2025 · similarity 0.90)
- Advancing Expert Specialization for Better MoE (NeurIPS 2025 · similarity 0.90)
- $\textit{Read-ME}$: Refactorizing LLMs as Router-Decoupled Mixture of Experts with System Co-Design (NeurIPS 2024 · similarity 0.90)
- ReMoE: Fully Differentiable Mixture-of-Experts with ReLU Routing (ICLR 2025 · similarity 0.90)
- dMoE: dLLMs with Learnable Block Experts (NeurIPS 2026 · similarity 0.89)
- Layerwise Recurrent Router for Mixture-of-Experts (ICLR 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.
