Research map / Architectures and efficiency / Mixture of experts
Optimizing LoRA Allocation of MoE with the Alignment of Topic Correlation
Hengyuan Xu, Wenjun Ke, Yao He, Jiajun Liu, Dong Nie, Peng Wang, Ziyu Shang, Zijie Xu
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 KL alignment of routing to topic prior (vs rigid topic routing) (ablation) | TopicLoRA · General | accuracy (%): 44.56 -> 52.98 | Improved |
| add multi-model cross-validation of topic prior (ablation) | TopicLoRA · Medical | accuracy (%): 42.82 -> 52.25 | Improved |
| add GNN graph representation learning on topic prior (ablation) | TopicLoRA · Law | accuracy (%): 36.64 -> 42.81 | Improved |
| TopicLoRA vs HydraLoRA on GSM8K | TopicLoRA · GSM8K | accuracy (%): 22.27 -> 26.16 | Improved |
| TopicLoRA vs HydraLoRA on Law | TopicLoRA · Law (MMLU) | accuracy (%): 42.18 -> 42.81 | No clear effect |
| TopicLoRA vs MoSLoRA on Medical | TopicLoRA · Medical (MMLU) | accuracy (%): 47.58 -> 48.25 | No clear effect |
| MoE-based LoRA (HydraLoRA) vs best MoE-free LoRA | HydraLoRA · MMLU | accuracy (%): 46.59 -> 47.22 | Improved |
| TopicLoRA vs HydraLoRA on multi-topic FlanV2/BBH | TopicLoRA · FlanV2 multi-topic (BBH) | accuracy (%): 41.5 -> 44.7 | Improved |
| increase number of LoRA experts from 2 to 8 (ablation) | TopicLoRA · MMLU subsets | accuracy | Improved |
| increase number of training topics from 2 to 6 (ablation) | TopicLoRA · mixed-topic training | accuracy (%): 22.89 -> 17.95 | Hurt |
| TopicLoRA vs HydraLoRA on HumanEval P@10 | TopicLoRA · HumanEval | pass@10 (%): 27.43 -> 26.28 | Hurt |
Closest papers by meaning
- From Experts to Sub-experts: Fine-grained Parameter-Efficient Fine-Tuning for MoE LLMs (NeurIPS 2026 · similarity 0.88)
- BAM! Just Like That: Simple and Efficient Parameter Upcycling for Mixture of Experts (NeurIPS 2024 · similarity 0.87)
- PT-MoE: An Efficient Finetuning Framework for Integrating Mixture-of-Experts into Prompt Tuning (NeurIPS 2025 · similarity 0.87)
- Pruning and Distilling Mixture-of-Experts into Dense Language Models (NeurIPS 2026 · similarity 0.87)
- Make LoRA Great Again: Boosting LoRA with Adaptive Singular Values and Mixture-of-Experts Optimization Alignment (ICML 2025 · similarity 0.86)
- FlyLoRA: Boosting Task Decoupling and Parameter Efficiency via Implicit Rank-Wise Mixture-of-Experts (NeurIPS 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.
