Research map / Architectures and efficiency / Mixture of experts
MoE LLM compression
Compresses mixture-of-experts LLMs through decomposition, delta decompression, and expert-level quantization.
15 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2024–2026) · 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.
2 papers in 2023–24 and 13 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 5.1% to 5.4%.
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
- 2024 · Pushing Mixture of Experts to the Limit: Extremely Parameter Efficient MoE for Instruction Tuning (ICLR 2024 · 11 citations)
- 2025 · MoE-SVD: Structured Mixture-of-Experts LLMs Compression via Singular Value Decomposition (ICML 2025 · 0 citations)
- 2026 · When Model Merging Breaks Routing: Training-Free Calibration for MoE (ICML 2026 · 0 citations)
Earliest papers here
- Pushing Mixture of Experts to the Limit: Extremely Parameter Efficient MoE for Instruction Tuning (ICLR 2024 · 11 citations)
- Merge, Then Compress: Demystify Efficient SMoE with Hints from Its Routing Policy (ICLR 2024 · 3 citations)
- MoE-SVD: Structured Mixture-of-Experts LLMs Compression via Singular Value Decomposition (ICML 2025 · 0 citations)
What has been tried
7 papers here report 91 measured changes (a component added, removed or swapped, compared with a baseline): helped 77%, no clear effect 3%, hurt 19%, mixed 1%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Increase MoV experts from 10 to 30 | P3 unseen tasks · median accuracy (Average) | 59.93 to 60.61 | Helped | Pushing Mixture of Experts to the Limit: Extremely Parameter Efficient MoE for Instruction Tuning (ICLR 2024 · 11 citations) |
| use an adaptive instead of uniform merging ratio | COPA · accuracy | 63.00 -> 68.00 | Helped | Merge, Then Compress: Demystify Efficient SMoE with Hints from Its Routing Policy (ICLR 2024 · 3 citations) |
| Add LoRA fine-tuning to MoE-SVD | 20% compression, average accuracy · accuracy | 0.58 -> 0.60 | Helped | MoE-SVD: Structured Mixture-of-Experts LLMs Compression via Singular Value Decomposition (ICML 2025 · 0 citations) |
| add a third layer to the 2-2 configuration | Winogrande · Accuracy | 0.87 | Helped | S'MoRE: Structural Mixture of Residual Experts for Parameter-Efficient LLM Fine-tuning (NeurIPS 2025 · 0 citations) |
| Combine discrete and continuous noise injection instead of injecting no noise | PIQA and SIQA (average), 5% level · accuracy (%) | 59.89 -> 62.66 | Helped | Noise-Robustness Through Noise: A Framework combining Asymmetric LoRA with Poisoning MoE (NeurIPS 2025 · 0 citations) |
| set soft prompt length to 40 tokens | MRQA in-domain · F1 | peak F1 60.66% | Helped | PT-MoE: An Efficient Finetuning Framework for Integrating Mixture-of-Experts into Prompt Tuning (NeurIPS 2025 · 0 citations) |
| Use Fisher-weighted merging instead of mean averaging | 40% compression · WikiText-2 perplexity | 7.66 -> 5.28 | Helped | Delta Decompression for MoE-based LLMs Compression (ICML 2025) |
What reviewers will expect
Datasets and baselines named in the 8 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- GSM8K · 3 of 8 papers
- WikiText-2 · 2 of 8 papers
- CommonsenseQA · 2 of 8 papers
- OpenBookQA · 2 of 8 papers
- SST-2 · 2 of 8 papers
- MRPC · 2 of 8 papers
- QNLI · 2 of 8 papers
Baselines
- LoRA · 4 of 8 papers
- HydraLoRA · 3 of 8 papers
- SparseGPT · 2 of 8 papers
- LoSparse · 2 of 8 papers
- ASVD · 2 of 8 papers
- MoE-I2 · 2 of 8 papers
- MC-SMoE · 2 of 8 papers
- AdaMoLE · 2 of 8 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 15 papers; from OpenAlex and the papers' affiliations).
- Hong Kong University of Science and Technology · 4 papers
- Cohere · 2 papers
- Imperial College London · 2 papers
- University of Birmingham · 2 papers
Use it
Download BibTeX of 15 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
- Mixture-of-experts LoRA (18 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).