Research map / Architectures and efficiency / Mixture of experts
Routing Manifold Alignment Improves Generalization of Mixture-of-Experts LLMs
Zhongyang Li, Ziyue Li, Tianyi Zhou
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 routing manifold regularization to router tuning (RoMA) | OLMoE-7B-A1B · MMLU | accuracy: 57.8 -> 69.0 | Improved |
| add RoMA | DeepSeekMoE-16B-A3B · MMLU | accuracy: 46.2 -> 56.8 | Improved |
| add RoMA | Qwen3-30B-A3B · MMLU | accuracy: 74.2 -> 78.8 | Improved |
| add RoMA to OLMoE on out-of-distribution math | OLMoE-7B-A1B · GSM8K | accuracy: 45.5 -> 49.4 | Improved |
| RoMA instead of C3PO test-time routing | OLMoE-7B-A1B · 8-benchmark average | accuracy: 75.7 -> 76.2 | No clear effect |
| regularize only last five router layers instead of all (ablation) | RoMA on OLMoE · 8-benchmark average | accuracy: 75.1 -> 76.2 | Improved |
| regularize last token only rather than first or middle token (ablation) | RoMA on OLMoE · 8-benchmark average | accuracy: 71.4 / 69.2 -> 76.2 | Improved |
| use k-NN successful neighbors instead of random neighbors (ablation) | RoMA on OLMoE · 8-benchmark average | accuracy: 67.8 -> 76.2 | Improved |
| use manifold regularization instead of L1, L2, or entropy (ablation) | RoMA on OLMoE · 8-benchmark average | accuracy: 71.5 -> 76.2 | Improved |
| increase k from 1 or 5 to 3 in k-NN neighbors (ablation) | RoMA on OLMoE · 8-benchmark average | accuracy | Improved |
| increase training data from 10% to 100% (ablation) | RoMA on OLMoE · 8-benchmark average | accuracy: 68.5 -> 76.2 | Improved |
| increase epsilon-ball radius beyond 0.5 to 0.7 (ablation) | RoMA on OLMoE · 8-benchmark average | accuracy | Hurt |
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 0.89)
- Your Mixture-of-Experts LLM Is Secretly an Embedding Model for Free (ICLR 2025 · similarity 0.89)
- Layerwise Recurrent Router for Mixture-of-Experts (ICLR 2025 · similarity 0.89)
- Tight Clusters Make Specialized Experts (ICLR 2025 · similarity 0.89)
- Routers Learn the Geometry of Their Experts: Geometric Coupling in Sparse Mixture-of-Experts (NeurIPS 2026 · 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.
