Research map / Architectures and efficiency / Mixture of experts
LoRA-Mixer: Coordinate Modular LoRA Experts Through Serial Attention Routing
Wenbing Li, Zikai Song, Hang Zhou, Junqing Yu, Yunyao Zhang, Wei Yang
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 RSL (entropy-regularized balance loss) instead of auxiliary loss, 1K routing data (ablation) | LoRA-Mixer · 7-task average (1K routing data) | average accuracy: 75.47 -> 76.80 | Improved |
| add RSL instead of auxiliary loss, 2K routing data (ablation) | LoRA-Mixer · 7-task average (2K routing data) | average accuracy: 77.29 -> 79.26 | Improved |
| add RSL instead of auxiliary loss, 4K routing data (ablation) | LoRA-Mixer · 7-task average (4K routing data) | average accuracy: 79.14 -> 78.77 | Hurt |
| add RSL instead of auxiliary loss, 10K routing data (ablation) | LoRA-Mixer · 7-task average (10K routing data) | average accuracy: 79.51 -> 79.94 | No clear effect |
| use RSL instead of AESL routing loss | LoRA-Mixer · HumanEval | Pass@1: 50.46 -> 57.32 | Improved |
| use RSL instead of GMoE routing loss | LoRA-Mixer · ARC-C | accuracy: 76.42 -> 83.24 | Improved |
| use alpha=0.1, lambda=0.05 instead of alpha=0.2, lambda=0.1 (ablation) | LoRA-Mixer · alpha/lambda sweep | performance (%): 75.8% -> 78.5% | Improved |
| use alpha=0.1, lambda=0.05 instead of alpha=0.05, lambda=0.01 (ablation) | LoRA-Mixer · alpha/lambda sweep | performance (%): 76.2% -> 78.5% | Improved |
| compose LoRAs with LoRA-Mixer over single-task LoRA | LoRA-Mixer · BoolQ | accuracy: 74.46 -> 79.37 | Improved |
| compose LoRAs with LoRA-Mixer over LoRA-LEGO on RTE | LoRA-Mixer · RTE (LLaMA2-7B) | accuracy: 71.85 -> 61.47 | Hurt |
| compose LoRAs with LoRA-Mixer over LoRA-LEGO on CoLA | LoRA-Mixer · CoLA (LLaMA2-7B) | accuracy: 55.48 -> 64.60 | Improved |
| compose LoRAs with LoRA-Mixer over single-task LoRA on GSM8K | LoRA-Mixer · GSM8K (Mistral-7B) | accuracy: 46.67 -> 46.48 | Hurt |
Closest papers by meaning
- FlyLoRA: Boosting Task Decoupling and Parameter Efficiency via Implicit Rank-Wise Mixture-of-Experts (NeurIPS 2025 · similarity 0.90)
- MeteoRA: Multiple-tasks Embedded LoRA for Large Language Models (ICLR 2025 · similarity 0.89)
- Merging LoRAs like Playing LEGO: Pushing the Modularity of LoRA to Extremes Through Rank-Wise Clustering (ICLR 2025 · similarity 0.89)
- LoRASuite: Efficient LoRA Adaptation Across Large Language Model Upgrades (NeurIPS 2025 · similarity 0.89)
- Beyond Higher Rank: Token-wise Input-Output Projections for Efficient Low-Rank Adaptation (NeurIPS 2025 · similarity 0.89)
- Beyond LoRA vs. Full Fine-Tuning: Gradient-Guided Optimizer Routing for LLM Adaptation (NeurIPS 2026 · 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.
