MEO

<h1 align="center">[EMNLP 2023] Merging Experts into One: Improving Computational Efficiency of Mixture of Experts</h1> <p align="center"> <a href="https://arxiv.org/abs/2310.09832"><img src="https://img.shields.io/badge/arXiv-2310.09832-b31b1b.svg" alt="arXiv"></a> <a href="https://aclanthology.org/2023.emnlp-main.907"><img src="https://img.shields.io/badge/EMNLP-2023-orange" alt="EMNLP 2023"></a> <a href="#-installation"><img src="https://img.shields.io/badge/Python-3.9%2B-green" alt="Python 3.9+"></a> </p> <p align="center"> Shwai He, Run-Ze Fan, Liang Ding, Li Shen, Tianyi Zhou, Dacheng Tao </p> <p align="center"> <a href="#-news">📰 News</a> • <a href="#-why-meo">🤔 Why</a> • <a href="#-core-idea">🧠 Core Idea</a> • <a href="#-quick-start">⚡ Quick Start</a> • <a href="#-installation">🛠️ Installation</a> </p>

📖 Introduction

This is the official implementation of the paper Merging Experts into One: Improving Computational Efficiency of Mixture of Experts, published at EMNLP 2023 Main Conference.

<p align="center"> <img src="Figures/MEO.png" width="860" alt="MEO overview"/> </p>

📰 News

• Dec 2023: Paper accepted at EMNLP 2023 Main Conference.
• Feb 2026: README updated with cleaner structure and quick-start workflow.

🤔 Why MEO

Sparse Mixture-of-Experts (MoE) improves model capacity and quality, but activating multiple experts usually increases computation cost.

MEO addresses this by merging multiple selected experts into one effective expert computation path, aiming to keep multi-expert benefits while reducing runtime overhead.

🧠 Core Idea

• ✅ Multi-expert selection is beneficial, but naive execution is expensive.
• ⚙️ MEO merges selected experts into one computation, reducing cost close to single-expert inference.
• 🔍 A token-level attention block is further introduced to improve token-level MEO efficiency and performance.

📁 Repository Structure

• tasks/text-classification/: GLUE and XNLI scripts.
• tasks/language-modeling/: CLM/MLM/PLM scripts.
• tasks/question-answering/: extractive and seq2seq QA scripts.
• tasks/summarization/: summarization scripts and task notes.
• transformers/: customized Transformers source used by this project.
• Figures/: project figures and result visualizations.

🛠️ Installation

1) 🧪 Create environment

conda create -n meo python=3.9 -y
conda activate meo

2) 📦 Install dependencies

pip install -r requirements.txt

⚡ Quick Start

You can run MEO-style experiments through task scripts below.

| Task | Dataset/Benchmark | Entry Script | |---|---|---| | Text Classification | GLUE | tasks/text-classification/run_glue.py | | Language Modeling | WikiText (CLM) | tasks/language-modeling/run_clm.py | | Question Answering | SQuAD (seq2seq) | tasks/question-answering/run_seq2seq_qa.py | | Summarization | XSum | tasks/summarization/run_summarization.py |

Example:

python tasks/text-classification/run_glue.py \
  --model_name_or_path bert-base-uncased \
  --task_name mrpc

📊 Results

<p align="center"> <img src="Figures/Results.png" width="680" alt="MEO results"/> </p>

From the paper, MEO provides substantial efficiency gains while preserving performance, for example:

• 📉 FLOPs reduced from 72.0G (vanilla MoE) to 28.6G (MEO).
• 🏆 On GLUE, token-level MEO reports 83.3% average score vs 82.6% for vanilla MoE in the reported setting.

For full setup details, please refer to the paper and scripts in this repository.

📚 Citation

@inproceedings{he-etal-2023-merging,
    title = "Merging Experts into One: Improving Computational Efficiency of Mixture of Experts",
    author = "He, Shwai  and
      Fan, Run-Ze  and
      Ding, Liang  and
      Shen, Li  and
      Zhou, Tianyi  and
      Tao, Dacheng",
    editor = "Bouamor, Houda  and
      Pino, Juan  and
      Bali, Kalika",
    booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2023",
    address = "Singapore",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.emnlp-main.907",
    doi = "10.18653/v1/2023.emnlp-main.907",
    pages = "14685--14691"
}

📬 Contact

For questions or collaboration, please contact: shwaihe@umd.edu.