A Survey on Mixture of Experts in LLMs: What the HKUST Repository Actually Ships
[TKDE'25] The official GitHub page for the survey paper "A Survey on Mixture of Experts in Large Language Models".
At a glance
- What is it?
- The withinmiaov repository is a paper list and taxonomy for a TKDE-accepted MoE survey, not a library. It is useful as a literature map, and useless as a dependency.
- Who is it for?
- Adopt this repository as a reading list if you are starting a MoE literature review or writing a related-work section, and skip it entirely if you need runnable code, since the repository contains no implementation, no releases and no installation path.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 29 days ago.
- What is it written in?
- GitHub does not report a main language for this repository.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What the repository contains, and what it does not
The name suggests a survey. The repository is the GitHub page for one. The survey itself lives on arXiv at 2407.06204, and the README states it has been accepted by TKDE. What you clone is a curated paper list, a taxonomy diagram, and a timeline image. There is no primary language listed for the repository, no releases have been retrieved, and the README contains no install command, no import statement and no script to run. If you arrived expecting a MoE implementation, a routing library or a benchmark harness, this is the wrong address. The intended reader is someone who needs the shape of the field: which MoE models exist, when they appeared, whether they are open or closed, and how routing, expert design and training work fit together.
The taxonomy and timeline images are the actual navigation layer
Two assets carry most of the orientation work. assets/moe_taxonomy.jpg lays out the survey's classification, and assets/MoE_Timeline_20250723.jpg is described in the README as a chronological overview of representative MoE models. The README explains the timeline's encoding rules precisely: models above the arrow are open-source, models below it are proprietary and closed-source, and colour marks the domain, with green for NLP, yellow for computer vision, pink for multimodal and cyan for recommender systems. That legend is the most useful sentence in the repository, because it tells you what the picture means without requiring you to read the paper first. A second image, assets/MoE_Timeline_Jan22.jpg, is labelled as the previous version from January 2025. Keeping both is a reasonable choice for anyone tracking how the field moved between snapshots.
How the paper list is organised and how to read it
The README's table of contents lists Taxonomy, Paper List (Organized Chronologically and Categorically), Contributors and Star History. The paper entries follow a consistent format: a title, a venue tag such as ICML 2026, CVPR 2026, ISCA 2026, AAAI 2026, ICLR 2026 or ArXiv 2026, a link to an arXiv identifier, and an explicit date. The list is sorted newest first, and the visible portion runs from early June 2026 back to mid-March 2026 before the supplied text is truncated. Entries span more than NLP: GEM applies graph-enhanced experts to dialogue state tracking, WaveMoE targets time series forecasting, PRISM and SMoES sit in vision and vision-language, and two ISCA 2026 papers address MoE serving and in-switch computing on multi-GPU systems. The venue tags are the practical filter. If you only care about systems work, ISCA 2026 marks the two infrastructure papers; if you care about interpretability, The Expert Strikes Back and Do Domain-specific Experts exist in MoE-based LLMs are the relevant entries.
Getting it running means cloning and reading
There is nothing to build. The only operation the material supports is fetching the repository and opening it, for example git clone https://github.com/withinmiaov/A-Survey-on-Mixture-of-Experts-in-LLMs, then reading the README and the images under assets/. The README also carries a PRs-Welcome badge and an explicit correction channel: readers are asked to email wcai738@connect.hkust-gz.edu.cn with mistakes or suggestions. That is the contribution path. There is no CI configuration described, no test suite, no dependency manifest and no versioning scheme in the supplied material, so treat the repository as a document rather than a maintained artifact. The last push timestamp is 2026-08-18, which sits after the newest paper entry visible in the README, so the list does appear to be receiving updates.
Where this repository is the wrong tool
The failure mode is category confusion. Someone searching for MoE tooling may land here because the repository name contains the phrase, then find no code, no releases and no installation instructions. The README does not offer a machine-readable export either: the paper list is Markdown prose, not BibTeX, not JSON, and not a dataset. If your pipeline needs to ingest citations programmatically, you will be writing a parser against a format that the maintainers have no stated obligation to keep stable. There is also a synchronisation risk between artifacts. The timeline image filename carries a 2025-07-23 date, the README labels the January 2025 image as the previous version, and the paper list extends into June 2026. Those three things are updated on different schedules, and nothing in the material guarantees the taxonomy diagram reflects the newest entries.
How it differs from Awesome lists and from code-first MoE projects
The closest comparison is the Awesome list pattern, which the README itself signals with an awesome.re badge. A typical Awesome list is a flat, community-maintained link dump with loose or no inclusion criteria. This repository is narrower and more opinionated: it is the companion page to a single peer-reviewed survey, so the entries are selected to serve that survey's taxonomy rather than to be exhaustive. The other comparison is a code-first project such as a MoE training framework, where the README documents configuration keys, parallelism settings and launch commands. That kind of repository answers how do I train or serve a MoE. This one answers what has been published about MoE and how the work clusters. The difference in approach is the deliverable: a curated bibliography with a taxonomy image versus an installable system. Neither substitutes for the other, and the README makes no claim that it does.
Maintenance cost, licence and what to verify
The repository is MIT licensed. For a bibliography that means the Markdown and the images can be reused with attribution, but the licence covers the repository contents, not the papers it links to, each of which carries its own terms. No legal advice implied; check the individual venues if you plan to redistribute figures. Maintenance cost on your side is low and mostly manual: you re-read the README when you need current entries, and you email corrections rather than opening a structured issue with a template, since none is described. The cost that does not go away is verification. The list is a pointer layer, so every claim you intend to make in your own writing has to be traced back to the arXiv identifier shown next to the entry. Start with the survey itself at arxiv.org/abs/2407.06204, then confirm the specific paper you are citing is still listed and that its venue tag matches the published record.
Editorial conclusion
Adopt this repository as a reading list if you are starting a MoE literature review or writing a related-work section, and skip it entirely if you need runnable code, since the repository contains no implementation, no releases and no installation path. Before relying on it, open the arXiv version at arxiv.org/abs/2407.06204 to confirm the taxonomy matches the current paper, check that the entries you plan to cite are still in the list, and note that the last push date is 2026-08-18 while the README's own timeline image is labelled January 2025 as the previous version, so the two artifacts are not guaranteed to be in sync.
Community notes