zhaoyang97/Paper-Notes: A 23,000-Note Index of AI Conference Papers

What Paper-Notes actually is, and who the notes are written for

Paper-Notes is a documentation site built with MkDocs, published at papernotes.org, whose content is a large set of short write-ups of conference papers. The README describes the format as reading a top-conference paper in five minutes, and the repository carries notes across ACL, CVPR, ICLR, AAAI, NeurIPS, ICCV, ICML and ECCV. The note count is stated as over 23,000, and individual conference pages carry their own totals, such as 5,342 for ICLR 2026 and 4,067 for CVPR 2026.

The audience is not the author of the paper, and it is not a reviewer. It is someone who has a proceedings list in front of them and needs to decide which twenty papers out of four thousand deserve a full read. The notes are written in Chinese, which narrows the audience further: an English-only reader gets the directory structure and the conference indexes, but not the note bodies. That is the single most consequential fact about the project, and it is easy to miss because the repository metadata and the README headings are partly in English.

A second thing to be clear about: this is an index and a summary layer, not a paper host. The README does not describe any mechanism for retrieving or mirroring the PDFs themselves. The value is the pointer plus the compressed idea, and the pointer matters because the companion repository zhaoyang97/papers-with-notes holds the full accepted-paper lists per conference.

How the repository is organised: conference, then research area, then one file per paper

The layout is a three-level tree under docs/. The first level is the conference, for example CVPR2026/ or ICML2026/. The second level is the research area, using the same folder names across conferences: 3d_vision/, llm_reasoning/, multimodal_vlm/, model_compression/ and so on. The third level is one markdown file per paper, named after a slug of the paper title.

Each level carries an index.md. The top-level docs/index.md is the site home and holds the full-site search entry point. A conference index aggregates that conference's notes by area. An area index aggregates the individual paper files. That means there are three ways to enter the content: search from the home page, browse a conference index, or browse an area index. The README does not describe the search implementation, only that the home page contains one.

The convention of reusing area folder names across every conference is the design decision that holds the whole thing together. It means a reader who follows llm_reasoning/ can move from ACL 2025 to ICLR 2026 without relearning the taxonomy. The cost is that the taxonomy is fixed at the repository level, and the README gives no route for a note that straddles two areas. The counts in the area table show how uneven the split is: image_generation/ holds 2,076 notes while federated_learning/ holds 19 and earth_science/ holds 19. Anyone who assumes the repository is balanced across AI subfields will be disappointed by the long tail.

Reading a note on the site, and what the math rendering warning means in practice

There is no install step for reading. The README points at the hosted site and gives one operational note: if a math formula renders incorrectly, refreshing the page usually resolves it. That is the only troubleshooting guidance the README offers, and it is worth taking literally. Formula-heavy papers are the ones where a broken render costs the most, and the stated remedy is a reload rather than a configuration change.

To read a specific paper, the path is conference index, then area index, then the paper file. The conference index URLs follow the pattern of the conference name, so the ECCV 2026 index sits under the ECCV2026 path and the ICLR 2026 index under ICLR2026. The companion paper lists live in a separate repository and are linked from the conference table.

If you want to run the site locally rather than use the hosted version, the repository ships mkdocs.yml at the top level along with an overrides/ directory and a hooks/ directory, which is the standard shape of a MkDocs project with a custom theme and build hooks. The README does not give a build command, a Python version, or a dependency file, so treat local builds as something you would have to work out from the MkDocs configuration rather than something the project documents.

Installing Paper-Notes locally with MkDocs

The README does not provide installation instructions, a requirements file, or a pinned MkDocs version, so the commands below follow the standard MkDocs workflow and the repository's own mkdocs.yml. Install MkDocs and start the development server from the repository root.

pip install mkdocs
mkdocs serve

The development server serves the site locally and rebuilds when a markdown file under docs/ changes, which is the useful mode if you are adding notes. To produce the static output instead, run the build command.

mkdocs build

That writes the rendered site to the directory named in mkdocs.yml. The repository also contains overrides/ and hooks/, which the MkDocs configuration references for theme customisation and build-time processing; if the build fails on a missing dependency, the failure will point at whatever those hooks import, and the README does not list it.

Adding a note means creating a markdown file under the right area folder and letting the area index pick it up. The README does not document whether the index files are generated by a script or maintained by hand, so before adding a large batch, check the scripts/ directory to see what is automated. Nothing in the README describes a validation step, so a malformed note will not be caught by the project.

Where the coverage is thin, and when this is the wrong tool

The most concrete limitation is the gap between the roadmap and the note counts. The roadmap lists NeurIPS 2026, EMNLP 2026, AAAI 2027, ICLR 2027 and CVPR 2027 with their notification and conference dates, but the coverage table does not include a note count for any of them. The roadmap is a schedule of when lists appear, not a commitment that notes will exist. A reader who arrives expecting NeurIPS 2026 notes will not find them in the coverage table.

The second limitation is the long tail in the area table. Nineteen notes under federated_learning/ and nineteen under earth_science/ are not a usable survey of either field. If your work sits in one of the small areas, the repository is a starting point at best, and you will exhaust it quickly.

The third is language. The notes are in Chinese. A team that needs English summaries cannot use the note bodies, only the structure.

The fourth is that a five-minute note is a compression, and compression loses the parts of a paper that matter for reproduction: training details, ablations, failure cases. Using these notes to decide what to read is reasonable. Using them as a substitute for reading, or as a citation source, is not. The README makes no claim about accuracy review, and there is no described review process.

How Paper-Notes differs from Papers with Code and from reading the proceedings directly

Papers with Code attaches code, datasets and benchmark leaderboards to papers. Its organising unit is the result: which method tops which benchmark, and where the implementation lives. Paper-Notes organises around the idea instead, with no leaderboard and no code index described in the README. If your question is which implementation to run, Papers with Code answers it and Paper-Notes does not. If your question is what the paper is arguing, the note is faster than the abstract because it is written to a fixed short format rather than to the author's own framing.

The other alternative is the proceedings themselves. Conference sites publish accepted-paper lists, and the companion repository zhaoyang97/papers-with-notes mirrors those lists as markdown files per conference. That route gives you the complete set with no editorial selection and no summary, which is the right choice when you already know the subfield and just need the full list. Paper-Notes adds value exactly where the list is too long to scan: 5,342 ICLR 2026 entries is not something most people will read title by title, and a per-area index with a one-paragraph idea attached is a real reduction in work.

Maintenance, licensing and what a fork commits you to

The repository is not archived and the last push was on 2026-09-06. The README lists releases through v1.5.0 on 2026-07-11, which added ECCV 2026 notes, and earlier releases that filled in ICLR 2026, CVPR 2026, ICML 2026 and ACL 2026. The pattern in those release notes is bulk completion of a conference rather than incremental addition, so the update cadence is tied to conference cycles. Between cycles, expect the repository to be quiet.

The licence field in the repository metadata reports NOASSERTION, but the README states the content is under CC BY-NC-SA 4.0, with attribution, non-commercial use, and share-alike. For a reader, that changes nothing. For anyone planning to reuse the notes, it changes a lot: non-commercial rules out bundling them into a paid product or an internal commercial knowledge base, and share-alike means a modified corpus has to carry the same licence. This is a description of the stated terms, not legal advice, and the metadata and README disagree, so confirm the terms before relying on them.

Upgrade cost is low if you only read the hosted site. If you fork and build locally, you inherit the MkDocs configuration, the overrides/ theme and the hooks/, and the README does not pin versions for any of them, so a future MkDocs release could break the build with no documented fallback.

Editorial conclusion

Adopt Paper-Notes if you read Chinese and want a fast orientation pass over ACL, CVPR, ICLR or NeurIPS proceedings before opening the PDFs, or if you want a MkDocs content tree to fork. Do not adopt it if you need notes in English, need coverage of every conference on the roadmap, or plan to reuse the content commercially, because the licence is CC BY-NC-SA 4.0. Verify first whether the specific conference folder and research-area folder you care about actually holds notes at the depth you need, and check the site at papernotes.org for the math rendering behaviour the README warns about.

Frequently asked questions

Official sources

1. Issues
2. Project website
3. README
4. zhaoyang97/Paper-Notes on GitHub