# An LLM multi-agent paper list that is one README, four images, and a taxonomy frozen in 2024 > LLM_MultiAgents_Survey_Papers keeps a five-stream bibliography of multi-agent LLM work. Its entries were last refreshed on different dates, one paper sits in two streams at once, and the Overview table it advertises is a PNG. **taichengguo/LLM_MultiAgents_Survey_Papers** — Large Language Model based Multi-Agents: A Survey of Progress and Challenges (In IJCAI 2024) - Repository: https://github.com/taichengguo/LLM_MultiAgents_Survey_Papers - Website: https://arxiv.org/abs/2402.01680 - Stars: 1,323 · Forks: 67 - Language: Unknown - License: not declared - Published: 2026-09-10 · Updated: 2026-09-10 · Language: en - Canonical page: https://hysenlabs.com/projects/taichengguo-llm-multiagents-survey-papers ## Five streams, and the repository is a README plus four images The whole project is five files. The root listing holds README.md, LLM-MA.png, image.png, overview.png, and trend.png. There is no source directory, no test suite, no CI configuration, no changelog, and no release. The repository has 1323 stars, 67 forks, and 2 open issues, and no GitHub releases exist for it at all. The most recent commit is dated 2026-08-21. The README divides its entries into five streams, announced in a News block stamped 2024/01: Multi-Agents Framework, Multi-Agents Orchestration and Efficiency, Multi-Agents for Problem Solving, Multi-Agents for World Simulation, and Multi-Agents Datasets and Benchmarks. The table of contents then splits two of those five into narrower subsections. Problem Solving expands into Software development, Embodied Agents, Science Team for Experiment Operations, Science Debate, and Database. World Simulation expands into Society, Game, Psychology, Economy, Recommender System, Policy Making, and Disease propagation Simulation. So calibrate the scale before reading. This is a reading list with a taxonomy attached, maintained by hand in a single Markdown file. Anything you want to do with it programmatically, from sorting by date to deduplicating across streams, you do yourself against the raw text. ## The architecture diagram and the Overview table are images, not text Two of the summary artifacts the README promises arrive as pictures. After the line Our summarized LLM-based Multi-Agents architecture is:, the page drops an empty centered block that carries an image. After The Overview table is as follows., another empty centered block does the same. Four PNG files at the root, LLM-MA.png among them, are what those blocks point at, alongside overview.png and trend.png. That matters more than decorative graphics would, because the Overview table is the artifact a newcomer opens first. A table inside an image cannot be searched, cannot be sorted, cannot be diffed when the classification changes, and cannot be read by a screen reader. When the page says more details can be seen in our paper, it is pointing at the only textual version of the classification that exists here. The survey itself is the anchor. The repository description names it as Large Language Model based Multi-Agents: A Survey of Progress and Challenges, published in IJCAI 2024 and hosted at arxiv.org/abs/2402.01680, and the News block repeats that address as where the survey is available. Everything else in the repository is a pointer into that paper or into other people's arxiv entries. ## A two-week update promise stamped 2024/02 beside entries stamped 2026/04 The News block opens with a cadence commitment: [2024/02] We will update our paper list every two weeks and include all the following papers in the next version of our paper. It then asks readers to contact the author in case papers have been missed. Nothing in the repository records whether that cadence was met. There is no changelog and no releases, so a biweekly schedule has no artifact to check itself against. What the file itself shows is the distance between the promise and the entries. The promise is stamped 2024/02. The newest entry anywhere in the streams is stamped 2026/04, and the newest commit is dated 2026-08-21. The list did keep moving after the promise was made, on a schedule the README never states. The other News line, stamped 2024/01, is the origin note: this repo is created to maintain LLM-based Multi-Agents papers, and the five streams were fixed at that moment. Both News lines are still the newest prose in the file even though the lists beneath them have grown by two years. A reader who treats the News block as a status page will draw the wrong conclusion about how current the rest of the file is. ## Orchestration reaches 2026/04 while Framework stops at 2024/03 The five streams are not kept to the same depth. Multi-Agents Framework holds thirteen entries running from 2024/03 down to 2023/03, and its newest item is a March 2024 paper on scaling laws for compound inference systems. Multi-Agents Orchestration and Efficiency holds thirteen entries as well, but they run from 2026/04 down to 2023/05. The Software development subsection under Problem Solving holds four entries, from 2024/02 down to 2023/10. So the newest Framework paper is more than two years older than the newest Orchestration paper, inside one file with one table of contents and one News block. Neither stream is marked as frozen. There is no per-stream date, no entry count, and no indication of which streams are still being fed and which stopped. Within each stream the ordering is consistent and newest first, which makes the gaps easier to find rather than harder to find. If you use the list to survey what exists, read it per stream rather than globally: the newest date anywhere in the file tells you about Orchestration and says nothing at all about Framework. ## L2MAC is listed twice and two titles collide on Is All You Need Deduplication is manual and it shows. L2MAC: Large Language Model Automatic Computer for Extensive Code Generation by Samuel Holt et al., arxiv 2310.02003, sits under Multi-Agents Framework at 2023/10 and again under Software development at the same 2023/10 with the same address. One work occupies two of the five streams at once. Titles are a second source of collision. The Framework stream opens with Are More LLM Calls All You Need? Towards Scaling Laws of Compound Inference Systems by Lingjiao Chen et al., arxiv 2403.02419. Eight months earlier the Orchestration stream carried More Agents Is All You Need by Junyou Li et al., arxiv 2402.05120. Two different papers, two author groups, two arxiv identifiers, near-identical titles, filed under different streams. Read the two lists as one flat bibliography, as the five-stream split invites, and you get a pair that looks duplicated without being duplicated. Every entry follows one line format: a bracketed year and month, then the title, then the author list, then a link. Two entries break the author rule. Achilles Heel of Distributed Multi-Agent Systems at 2025/04 carries no author at all, and crewAI at 2023/11 credits joaomdmoura et al., a GitHub handle where every other entry names people. ## Thirty entries and only two links that do not point at a paper Almost every entry leads to one place. Of the thirty entries across the Framework, Orchestration, and Software development lists, twenty-eight resolve to an arxiv.org abstract page and nothing more. There is no link to a repository, no dataset, no leaderboard, and no evaluation harness, partly because there is nothing in this repository to link to. Two entries break the pattern, and both help precisely because they do. CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery by Ao Qu et al. at 2026/04 carries both a paper link and a code link, pointing at a project under the Human-Agent-Society organization. The crewAI entry at 2023/11 carries no paper at all, only a repo link, which makes it the one entry in the list that names a tool rather than a publication. That ratio is the practical limit of the resource. The list tells you what to read and who wrote it. Getting from a title to something you can run is, in twenty-eight cases out of thirty, your own next step. ## The citation request sits inside an HTML comment The README asks to be cited and then hides the request. A Citation heading and the line Should you find value in our paper and this repository, we would be most grateful if you could cite our paper: are both wrapped in an HTML comment, so a rendered page shows nothing there. The table of contents also promises two sections a reader never reaches from the top of the page: Contributing and Contact. No file sits behind Contributing in the root listing. The Software development subsection under Problem Solving runs from a February 2024 paper on large language models serving as data analysts, through XUAT-Copilot, AgentCoder, and L2MAC, and stops there. None of that makes the entries wrong. It makes the file a hand-maintained artifact rather than a maintained dataset, and that distinction is the one to carry into any use of it. Read it as a curated reading list with an uneven refresh rate, a stream split chosen in early 2024, and a classification chart that exists only as a picture. ## Conclusion Treat this repository as a reading list with a fixed taxonomy and an uneven refresh rate, not as a dataset. It is the right starting point if you want the IJCAI 2024 survey's classification of LLM multi-agent work plus pointers added over the two years after it. It is the wrong tool if you need coverage you can filter, a record of what changed, or assurance that a stream is current. Before building on it, read the newest date in each of the five streams separately, check that a paper is the one you meant when two titles look alike, and remember the Overview table exists only as an image. ## FAQ ### What does the LLM_MultiAgents_Survey_Papers repository actually contain? A README and four PNG images at the root, with no source code and no releases. The README groups entries into five streams: Multi-Agents Framework, Multi-Agents Orchestration and Efficiency, Multi-Agents for Problem Solving, Multi-Agents for World Simulation, and Multi-Agents Datasets and Benchmarks. ### What language and license does the LLM multi-agent survey papers repository use? Neither is recorded. GitHub shows no primary language and no license for it, and the root holds only README.md and four images with no license file among them. The survey it points to is arXiv 2402.01680, published in IJCAI 2024. ### How often is the LLM-based multi-agents paper list updated? The News section promises an update every two weeks in an entry stamped 2024/02, and nothing in the repository records whether that held. The newest commit is dated 2026-08-21 and the newest entry anywhere is stamped 2026/04, while the Framework stream's newest entry is 2024/03. ### Can I filter or export the paper list from the LLM_MultiAgents_Survey_Papers repo? No such tooling is present. The root holds README.md and four images, there are no releases, and the entries are Markdown list items whose links go to arxiv.org. The architecture summary and the Overview table are delivered as PNG files rather than as text. ## Sources - [Issues](https://github.com/taichengguo/LLM_MultiAgents_Survey_Papers/issues) - [Project website](https://arxiv.org/abs/2402.01680) - [README](https://github.com/taichengguo/LLM_MultiAgents_Survey_Papers/blob/main/README.md) - [taichengguo/LLM_MultiAgents_Survey_Papers on GitHub](https://github.com/taichengguo/LLM_MultiAgents_Survey_Papers) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/taichengguo-llm-multiagents-survey-papers