Research map / Architectures and efficiency / Mixture of experts
Mixture-of-experts LoRA
Combines low-rank adapters with expert routing for multi-task and parameter-efficient adaptation.
18 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Mixture of experts, keyword cluster “lora · rank · tuning”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
3 papers in 2023–24 and 14 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 7.7% to 5.9%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2017 · Expert Gate: Lifelong Learning With a Network of Experts (CVPR 2017 · 613 citations)
- 2024 · Multi-Task Dense Prediction via Mixture of Low-Rank Experts (CVPR 2024 · 50 citations)
- 2025 · SPMTrack: Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking (CVPR 2025 · 16 citations)
- 2026 · D2MoRA: Diversity-Regulated Asymmetric MoE-LoRA Decomposition for Efficient Multi-Task Adaptation (AAAI 2026 · 2 citations)
Earliest papers here
- Expert Gate: Lifelong Learning With a Network of Experts (CVPR 2017 · 613 citations)
- Multi-Task Dense Prediction via Mixture of Low-Rank Experts (CVPR 2024 · 50 citations)
- Merging Multi-Task Models via Weight-Ensembling Mixture of Experts (ICML 2024 · 2 citations)
What has been tried
11 papers here report 136 measured changes (a component added, removed or swapped, compared with a baseline): helped 73%, no clear effect 6%, hurt 19%, mixed 2%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add the task-sharing generic convolution path to the low-rank expert path | PASCAL-Context · MTL ∆m | -1.71 -> -0.88 | Helped | Multi-Task Dense Prediction via Mixture of Low-Rank Experts (CVPR 2024 · 50 citations) |
| add TMoE instead of using LoRA for parameter-efficient fine-tuning | LaSOT test · AUC (%) | 72.9 -> 73.7 | Helped | SPMTrack: Spatio-Temporal Parameter-Efficient Fine-Tuning with Mixture of Experts for Scalable Visual Tracking (CVPR 2025 · 16 citations) |
| Replace weight averaging with other tested merging techniques while retaining Best-in-TIME initialization and deployment | FoMo-in-Flux · knowledge accumulation | No clear effect | How to Merge Your Multimodal Models Over Time? (CVPR 2025 · 3 citations) | |
| use a two-layer rather than zero-layer router | Seen tasks (six-task average) · top-1 accuracy | 73.5 -> 86.5 | Helped | Merging Multi-Task Models via Weight-Ensembling Mixture of Experts (ICML 2024 · 2 citations) |
| Add MoEfied LoRA | Multi-task FGVC · Mean top-1 accuracy | 87.39 -> 90.27 | Helped | Transforming Vision Transformer: Towards Efficient Multi-Task Asynchronous Learner (NeurIPS 2024 · 2 citations) |
| Add load balancing | MMLU · accuracy | 37.56 -> 40.88 | Helped | FlyLoRA: Boosting Task Decoupling and Parameter Efficiency via Implicit Rank-Wise Mixture-of-Experts (NeurIPS 2025 · 1 citation) |
| Use non-linear rather than linear reparameterization | Multiple domains · performance | Helped | RepLoRA: Reparameterizing Low-rank Adaptation via the Perspective of Mixture of Experts (ICML 2025 · 0 citations) | |
| use gate-based forward revision without Riemannian preconditioning | ScienceQA · training loss | No clear effect | A Stronger Mixture of Low-Rank Experts for Fine-Tuning Foundation Models (ICML 2025 · 0 citations) | |
| Use amplitude-aware instead of random masking during pretraining | TUEV · Balanced Accuracy | 0.6320 -> 0.6845; 67.35 -> 67.61 | Helped | NeurIPT: Foundation Model for Neural Interfaces (NeurIPS 2025 · 0 citations) |
| add the orthogonal adapter (increase L from 0 to 8) | CSR-MTL · Overall (%) | 84.98 -> 85.11 | Helped | MoORE: SVD-based Model MoE-ization for Conflict- and Oblivion-Resistant Multi-Task Adaptation (NeurIPS 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 7 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ScienceQA · 2 of 7 papers
Baselines
- Full fine-tuning · 2 of 7 papers
- LoRA · 2 of 7 papers
- AdaLoRA · 2 of 7 papers
- HydraLoRA · 2 of 7 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 17 papers; from OpenAlex and the papers' affiliations).
- Beihang University · 2 papers
- Tsinghua University · 2 papers
Use it
Download BibTeX of 18 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- MoE LLM compression (15 papers)
Elsewhere in the topic
- Expert routing and specialization (150 papers)
- Efficient MoE language model systems (72 papers)
- Sparse MoE as regularization and attention (45 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).