Research map / Architectures and efficiency / Mixture of experts
LiME: Lightweight Mixture of Experts for Efficient Multimodal Multi-task Learning
Md Kowsher, Haris Mansoor, Nusrat Jahan Prottasha, Ozlem Garibay, Victor Zhu, Zhengping Ji, Chen Chen
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 separate per-expert adapters with a shared PEFT module plus lightweight expert vectors | LiME-LoRA vs MoELoRA · GLUE efficiency | trainable parameters: 1.97M -> 0.52M | Improved |
| replace separate per-expert adapters with a shared PEFT module plus expert vectors | LiME-DoRA vs MoE-DoRA · GLUE efficiency | training time (min): 50.2 -> 35.5 | Improved |
| add LiME modulation and zero-parameter routing to base PEFT methods | LoRA, DoRA, LoRA-FA, SliceFine, Prompt Tuning · MMT-47 categories | average score: Vision Benchmark LoRA 77.02 -> LiME-LoRA 78.01 | Improved |
| use zero-parameter routing instead of a learned router (ablation) | LiME · text and vision tasks | accuracy | No clear effect |
| choose routing feature dimensions (leading, central, trailing, random E) (ablation) | LiME routing · Figure 4a | accuracy | No clear effect |
| mix frozen and PEFT outputs for routing (gamma_r around 0.7) instead of either alone (ablation) | LiME routing · text and vision tasks | accuracy | Improved |
| use Auto Top-K relative threshold theta 0.7 instead of fixed top-k (ablation) | LiME · GLUE SST-2 | accuracy: 92.08 -> 92.43 (k=3 vs theta=0.7) | Improved |
| activate all four experts (fixed k = 4) instead of k = 3 (ablation) | LiME fixed Top-K · GLUE SST-2 | accuracy: 92.08 -> 91.94 | Hurt |
| set relative threshold too strict (theta 0.8 instead of 0.7) (ablation) | LiME Auto Top-K · GLUE STS-B | accuracy: 81.87 -> 81.54 | Hurt |
| add load balancing losses (coefficient about 1e-2 to 0.1) versus none (ablation) | LiME · GLUE | accuracy | Improved |
| over-strong load balancing (coefficient above 1.0) (ablation) | LiME · GLUE | accuracy | Hurt |
| increase number of experts from 1 to 4-5, then beyond 6 (ablation) | LiME · GLUE | accuracy | Mixed |
Closest papers by meaning
- From Experts to Sub-experts: Fine-grained Parameter-Efficient Fine-Tuning for MoE LLMs (NeurIPS 2026 · similarity 0.88)
- R2-T2: Re-Routing in Test-Time for Multimodal Mixture-of-Experts (ICML 2025 · similarity 0.86)
- Multi-Task Dense Prediction via Mixture of Low-Rank Experts (CVPR 2024 · similarity 0.86)
- FLAME: Adaptive Mixture-of-Experts for Continual Multimodal Multi-Task Learning (NeurIPS 2026 · similarity 0.86)
- FlyLoRA: Boosting Task Decoupling and Parameter Efficiency via Implicit Rank-Wise Mixture-of-Experts (NeurIPS 2025 · similarity 0.86)
- Multilinear Mixture of Experts: Scalable Expert Specialization through Factorization (NeurIPS 2024 · 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.
