Top-AI-Conferences-Paper-with-Code: a Markdown index of conference papers that ship code

The gap this repository fills: papers whose code you can actually find

Conference papers are short. The README states the motivation plainly: a paper's length rarely covers every implementation detail, and accompanying open source code provides more reproducibility information and makes follow-up work easier. Anyone who has tried to reimplement a result from a proceedings PDF knows the failure mode. The paper describes the method, the appendix lists hyperparameters, and the actual training script lives in a GitHub repository that the paper never links or links under a different name than the project uses.

This repository addresses a narrower problem than "find me papers". It answers "which papers from this conference year published code, and where is it". The unit of organisation is the conference and the year, not the topic or the method. That makes it useful for a specific kind of task: surveying what was released at a venue you already care about, rather than discovering a venue for a topic you already care about. The README names the covered venues as ACL, EMNLP, NAACL, COLING, AAAI, IJCAI, ICLR, NeurIPS and ICML, and the topics list adds computer-vision, deep-learning, machine-learning, natural-language-processing, PyTorch and TensorFlow, so the scope is wider than the MLNLP name suggests.

The audience is narrow and identifiable. A graduate student starting a literature review for a specific venue. A research engineer checking whether the baseline they want to compare against has a public implementation. A lab that runs internal reading groups and wants a per-conference list to assign. None of these people need an API. They need a file they can open, search with Ctrl+F, and share as a link.

How the repository is laid out: one Markdown file per conference per year

There is no build system, no package, no server. The top-level repository entries are the conference directories (AAAI/, ACL/, COLING/, EMNLP/, ICLR/, ICML/, IJCAI/, NAACL/, NeurIPS/), plus README.md, pics/ and .gitignore. Inside each conference directory the structure is year-first: ACL/2019/ACL2019.md, ICLR/2025/ICLR2025.md, and so on. The README's support table links directly to those files, with a dash in the cells where no file exists.

That layout is the entire architecture, and it has consequences worth naming. Because the content is Markdown, it renders on GitHub, clones in seconds, and diffs cleanly when someone adds a paper. Because it is organised by year, a new conference edition means a new file rather than an edit to an existing one, which keeps contributions from colliding. Because the README table is the only index, a missing link in the table is the only signal that a year is uncovered. There is no generated index, no search endpoint, and no schema. If you want to query this data programmatically you would have to parse the Markdown yourself, and the repository does not document a format for the entries, so any parser you write is coupled to whatever conventions the contributors happened to follow.

The update log shows this is maintained by hand. The 2026-07-2 entry describes reorganising and updating the README and adding a 2026 entry point with Chinese notes; the 2026-07-1 entry describes updating ICLR coverage and adding ICLR2025 and ICLR2026 files. The last push to the repository was on 2026-07-02. There are no releases.

Opening the files: what a first use looks like

There is nothing to install. The README documents no package, no CLI and no configuration. The project is a set of Markdown files, so the first real use is reading the file for the venue and year you care about. The README's support table is the map: find your conference row, then the year column, and follow the link. For example, the ICLR row links to ICLR/2025/ICLR2025.md and ICLR/2026/ICLR2026.md, which the update log says were added on 2026-07-1.

Those links point at files on GitHub, and the same paths work in a local copy of the repository. The directory names are the ones listed in the repository root, so the path pattern is consistent across venues: a conference directory, a year directory, and a Markdown file named after the conference and year. If you are checking coverage before you start reading, the file paths themselves tell you what exists, because a year with no file simply has no path.

Coverage gaps are the real limitation, and the README table admits them

The support table is unusually honest about what is missing, which is a point in the project's favour and also the clearest statement of its limits. NAACL has 2019, 2021, 2022, 2024 and 2025, with dashes for 2020 and 2023. COLING jumps from 2018 to 2020, then 2022, 2024 and 2025, skipping 2021 and 2023. AAAI and IJCAI both stop at 2025 in the visible table. ACL and EMNLP are covered from 2019 through 2025 with 2026 left open. Anyone who needs a complete run of a venue will not get it here, and no amount of browsing will produce a file that does not exist.

The second limitation is that inclusion depends on someone noticing a paper and adding it. The README invites contributions through issues and pull requests, and the badges mark the status as building and PRs as welcome. That is a volunteer model. It means coverage within a single file is not guaranteed to be exhaustive, and it means the freshness of any given file depends on when a contributor last touched it. A file can exist and still be months behind the conference's full program.

The third limitation is the absence of any machine-readable structure. There is no JSON, no CSV, no front matter documented in the README. If your workflow needs to filter by task, dataset, or framework, this repository will not do it. You would be reading prose.

Finally, the licence is not stated anywhere in the README or the repository description. That is a genuine blocker for anyone who wants to redistribute the collected content, and it is worth resolving before you build anything on top of it.

How it differs from a searchable paper database and from reading the proceedings yourself

A database-backed paper index is the obvious comparison, and the difference is not quality but shape. A database gives you per-paper pages, task and benchmark taxonomies, and links between results and leaderboards. You can filter, sort and follow a task across years. This repository is a set of static files organised by conference and year. It has no database, no filters and no per-paper pages beyond whatever a contributor typed into the Markdown.

The practical consequence is that the two approaches answer different questions. If you want to know the state of the art on a specific task, a database is built for that and this repository is not. If you want to know what shipped code at ICLR 2026, this repository gives you a single file to read, which is faster than paging through a web interface and works offline after a clone. The trade is coverage and structure for a flat, greppable, version-controlled list.

The second alternative is simply reading the conference proceedings or the accepted-paper list and checking each paper for a code link. That is the ground truth and it will always be more complete than a curated index. It is also slow, and it is the work this repository exists to save you. The honest framing is that this project is a shortcut, and like any shortcut it is only as good as the person who laid it out.

Maintenance, contributions and what the licence silence means

The repository is not archived, and the last push was on 2026-07-02. The update log records two consecutive days of work in July 2026, reorganising the README and adding ICLR2025 and ICLR2026 files. There are no releases, which is consistent with a content repository: versioning happens through commits, and the README carries a v0.1.0 badge that does not correspond to any published artifact.

Upgrade cost is close to zero in the software sense. There is no dependency to bump and no migration to run. Updating means pulling the branch and re-reading the files you care about, and because the content is Markdown, a pull will show you exactly which lines changed. The cost that does exist is editorial: if you fork the repository to add your own venue or year, you are maintaining that file by hand, and nothing in the repository will tell you when the upstream conference publishes a new program.

On licensing, the README gives no licence identifier for the repository. The papers and code it links to carry their own licences, which are entirely separate and which you must check at each linked repository before reusing anything. Redistributing the index itself, or bundling it into a product, is a question the repository does not answer, and that is a gap rather than a permission.

Editorial conclusion

Adopt it if you need a fast way to scan which papers from a given conference year published code, and you are willing to tolerate gaps in the years the maintainers have not covered. Do not adopt it as a substitute for a searchable paper database if you need complete coverage, filtering, or any kind of programmatic query. Before relying on it, open the file for the specific conference and year you care about and check whether that cell in the support table is filled in or left as a dash.

Frequently asked questions

Official sources

1. Issues
2. MLNLP-World/Top-AI-Conferences-Paper-with-Code on GitHub
3. README