Towards Efficient Multimodal Large Language Models: A Survey on Token Compression

📢 Contributions Welcome

We appreciate contributions that help improve this repository and the accompanying paper. Please feel free to submit a pull request to:

1. Add a missing or relevant paper.
2. Propose a more suitable category or tag.
3. Update or correct information (links, metadata, status).
4. Addressing Potential Issues in Benchmarking.
5. Request clarification or report an issue.

Thank you — every suggestion helps make this resource more useful.

⭐ If you find this repo useful, please give us a star :)

✒️ Table of Contents

• News: Latest Updates, News, and Announcements.
• About: Overview and Objectives.
• Citation: If you find this helpful, please consider citing us.
• Tag Description: Brief Explanation of Tags in Paper Table.
• Paper Table: Paper Index (by Year, Descending).
• Benchmark: An overview of our proposed benchmark for MLLM token compression .

🔥 News

• [2025.12.18] We've released the first version (v1.0) of the survey, which can be downloaded here.
• [2025.11.26] We've released the repository!

☀️ About

Multimodal Large Language Models (MLLMs) are rapidly expanding their capabilities, but high-resolution images and long videos create extremely long visual-token streams that dramatically increase compute, memory, and latency requirements. This repository accompanies our survey on Towards Efficient Multimodal Large Language Models: A Survey on Token Compression (TechRxiv) to help researchers and practitioners navigate this field.

Motivation. Token compression reduces the number of visual tokens processed by MLLMs while preserving critical cross-modal semantics, enabling more efficient training and faster inference without large accuracy regressions. The field is fragmented across encoders, projectors, and LLM-side techniques; a centralized, searchable resource is needed.

Target audience. Researchers, implementers, and system designers working on multimodal models, retrieval, efficient vision-language pipelines, and deployment at scale.

What this repo provides.

• A curated, searchable, chronologically organized paper index.
• Short annotated entries with metadata (method family, compression ratio, retrain vs plugin, modality).
• Links to code, checkpoints, and reproducibility notes where available.
• An overview of our proposed evaluation suite for MLLM token compression and benchmarks.
• Contribution guidelines and templates for adding papers or methods.

Feel free to browse the table, open issues, or contribute entries to help grow a rigorous, practical ecosystem for efficient multimodal modeling.

📌 Citation

If you find our paper or this resource helpful, please consider cite:

@misc{yao2026towards,
  title        = {Towards Efficient Multimodal Large Language Models: A Survey on Token Compression},
  author       = {Yao, Linli and Xing, Long and Shi, Yang and Li, Sida and Liu, Yuanxin and
                  Dong, Yuhao and Zhang, Yi-Fan and Li, Lei and Dong, Qingxiu and Dong, Xiaoyi and
                  Huang, Qidong and Wang, Haotian and Wu, Feng and Zhang, Yuanxing and Wan, Pengfei and
                  Lin, Zhouchen and Sun, Xu},
  year         = {2026},
  month        = jan,
  howpublished = {TechRxiv},
  doi          = {10.36227/techrxiv.176823010.07236701/v1},
  url          = {https://doi.org/10.36227/techrxiv.176823010.07236701/v1}
}

📋 Tag Description

• red for arXiv papers
• blue for conference/journal papers
• white for GitHub repositories
• purple for modality
• cyan for compression position
• brightgreen for whether it is text query-based
• lightgrey for compression methods: merge or pruning
• yellow for usage mode: re-train or plug-in
• orange for acceleration stage: Train stage or Inference Stage
• pink for compression ratio: fix or dynamic
• yellowgreen for usage stage

📚 Paper Table