Research map / Architectures and efficiency / Mixture of experts
LD-MoLE: Learnable Dynamic Routing for Mixture of LoRA Experts
Yuan Zhuang, Yi Shen, Yuexin Bian, Qing Su, Shihao Ji, Yuanyuan Shi, Fei Miao
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 |
|---|---|---|---|
| replace TopK routing with learnable dynamic Sparsegen routing | MoLA(8888) · 9 benchmarks, Qwen3-1.7B | average score: 79.24 -> 81.63 | Improved |
| replace TopK routing with learnable dynamic Sparsegen routing | MoLA(8888) · 9 benchmarks, Llama3.2-3B | average score: 78.18 -> 82.05 | Improved |
| replace ReLU routing with Sparsegen routing and predicted lambda | ReMoLE · 9 benchmarks, Llama3.1-8B | average score: 84.47 -> 85.18 | Improved |
| replace ReLU routing with Sparsegen routing and predicted lambda | ReMoLE · 9 benchmarks, Qwen3-1.7B | average score: 81.02 -> 81.63 | Improved |
| prune experts per layer (2468) instead of uniform 8 | MoLA · 9 benchmarks, Llama3.2-3B | average score: 78.18 -> 78.72 | Improved |
| use predicted lambda instead of fixed lambda 0.5 (ablation) | LD-MoLE · 5 datasets, Qwen3-1.7B | average score: 83.66 -> 84.56 | Improved |
| use predicted lambda instead of fixed lambda -10.0 (ablation) | LD-MoLE · 5 datasets, Qwen3-1.7B | average score: 83.35 -> 84.56 | Improved |
| add sparsity loss with coefficient 1.0 (ablation) | LD-MoLE · 5 datasets, Qwen3-1.7B | average score: 84.56 -> 83.11 | Hurt |
| add sparsity loss with coefficient 0.01 (ablation) | LD-MoLE · 5 datasets, Qwen3-1.7B | average score: 84.56 -> 84.02 | Hurt |
| add sparsity loss with coefficient 0.01 (beta) (ablation) | LD-MoLE · Qwen3-1.7B | MFLOPs: 106 -> 100 | Improved |
| add sparsity loss with beta 1.0 (ablation) | LD-MoLE · Qwen3-1.7B | MFLOPs: 106 -> 83 | Improved |
| use shared MLP instead of local linear lambda predictor (ablation) | LD-MoLE · 8 tasks, Qwen3-1.7B | average score: 84.65 -> 85.05 | Improved |
Closest papers by meaning
- Ada-K Routing: Boosting the Efficiency of MoE-based LLMs (ICLR 2025 · similarity 0.90)
- Routing Experts: Learning to Route Dynamic Experts in Existing Multi-modal Large Language Models (ICLR 2025 · similarity 0.90)
- Layerwise Recurrent Router for Mixture-of-Experts (ICLR 2025 · similarity 0.90)
- Advancing Expert Specialization for Better MoE (NeurIPS 2025 · similarity 0.88)
- From Experts to Sub-experts: Fine-grained Parameter-Efficient Fine-Tuning for MoE LLMs (NeurIPS 2026 · similarity 0.88)
- $\textit{Read-ME}$: Refactorizing LLMs as Router-Decoupled Mixture of Experts with System Co-Design (NeurIPS 2024 · similarity 0.87)
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.
