Paper-Notes: A Markdown Corpus of Conference Paper Summaries

The gap between a conference proceedings and a reading list

A conference like CVPR 2026 accepts thousands of papers. The README lists 4,067 notes for that venue alone, and 5,342 for ICLR 2026. Nobody reads that. The practical problem is triage: deciding which of several thousand titles deserves an hour of your time. Paper-Notes addresses that by attaching a short written summary to each paper and organising the summaries by venue and by research area, so you can scan a folder such as `llm_reasoning/` or `multimodal_vlm/` instead of a proceedings PDF. The intended reader is an engineer or a student who wants the core idea of a paper in roughly five minutes, which is the claim the repository makes in its own description. It is not a reimplementation, not a benchmark harness, and not a library you import. The deliverable is prose.

How the corpus is laid out on disk

The README gives the directory structure directly. Everything lives under `docs/`, with one folder per venue (`CVPR2026/`, `ICML2026/`, `ACL2026/`, `ICLR2026/`, `AAAI2026/`, `NeurIPS2025/`, `ICCV2025/`, `ACL2025/`, `ICML2025/`, `CVPR2025/`, `ECCV2024/`). Inside a venue folder sit topic folders such as `3d_vision/`, and inside those sit the individual notes as Markdown files named after the paper slug. Each level carries an `index.md`: the root `docs/index.md` is described as the site home with full-site search, a venue `index.md` aggregates by area, and a topic `index.md` lists the notes in that folder. That is a plain static-site pattern. The consequence is that the repository is greppable offline. You can clone it and run a text search across every note without a network connection, which is the main reason to prefer the repository over the hosted site. The Python language tag on the repository most likely reflects tooling around the content rather than the notes themselves, and the supplied material does not describe that tooling, so treat the Python classification as unconfirmed in purpose.

What the coverage table actually tells you

The coverage table is the most useful thing in the README, because it separates two different numbers that are easy to confuse. One column is the note count per venue; the other links to a separate repository, `zhaoyang97/papers-with-notes`, holding a file like `CVPR2026-Papers.md` described as the complete paper list. So the note count and the accepted-paper count are not the same figure, and the README does not state what fraction of each venue is covered. For ICLR 2026 the release notes claim the full set of roughly 5,000 papers was completed in v1.4.0, and for CVPR 2026 roughly 4,000 in v1.3.0. Those are the author's own figures. The topic table is the other half of the picture: 54 folders are listed, from `3d_vision/` at 2,010 notes down to `anomaly_detection/` at 14. The distribution is uneven, and areas with small counts are thin enough that you should check the folder before assuming coverage.

Reading it online versus cloning it

There is no installation step. The README points to https://papernotes.org/ as the reading surface and offers no build instructions, no dependency list and no local server command. If you want the corpus on your own machine, the only route the material supports is cloning the repository and reading the Markdown under `docs/` directly, or rendering it with whatever static site generator you prefer. The README does not name one. One operational note appears in the README itself: if mathematical formulas render incorrectly, refreshing the page usually fixes it. That is a client-side rendering quirk on the hosted site, and it is worth knowing before you conclude that a note is broken. Because there are no releases retrieved for this repository, the version numbers in the release notes (v1.0.0 through v1.5.0) are content milestones described in the README, not downloadable artifacts. Do not go looking for a tagged tarball.

The licence is the constraint that decides most use cases

The README states the content is licensed CC BY-NC-SA 4.0, with three terms spelled out: attribution, non-commercial use, and share-alike for modifications. The repository's machine-readable licence field reads NOASSERTION, which means GitHub could not map the repository to a standard licence file. Those two signals point the same direction but are not identical, and the discrepancy is worth noting rather than resolving. Practically, the non-commercial clause rules out the obvious enterprise uses: bundling notes into a paid product, feeding them into a commercial retrieval index, or using them as training data for a commercial model. The share-alike clause means a derived and redistributed version has to carry the same terms. This is not legal advice, and if the distinction matters to you, the licence deed linked from the README is the document to read. The safe reading of the material is that this is a non-commercial reference corpus.

Where the model breaks down

The failure mode is quality variance, and the structure of the project makes it structural rather than incidental. A single maintainer producing thousands of notes per venue, on the schedule implied by the roadmap (ECCV 2026 notes after the July 2026 list publication, NeurIPS 2026 after September 2026, and so on), cannot give each paper the same attention. The README gives no per-note review process, no indication of who writes the notes, and no correction mechanism beyond opening an issue. There is no stated policy on what happens when a note is wrong. That matters more here than in a normal documentation project, because a reader who skims a five-minute summary and skips the paper has no independent check on the summary's accuracy. The other limitation is scope: the corpus covers conferences, so a paper that appeared only as an arXiv preprint, or at a workshop, or in a journal, is out of scope by construction. If your reading list is not conference-shaped, this is the wrong index.

Compared with Semantic Scholar and the papers-with-notes list

The natural alternative for the triage problem is Semantic Scholar or an equivalent scholarly search index. The difference is in what gets produced. A search index gives you metadata: title, abstract, authors, citation links, and increasingly an automatically generated summary. Paper-Notes gives you a human-written note organised into a topic folder alongside its neighbours. The trade-off runs both ways. The search index covers far more papers and is not limited to conference cycles, but it will not tell you that this paper sits next to forty other LLM reasoning papers from the same venue. Paper-Notes is bounded by its venue list, but the folder structure is itself information. A second, smaller alternative sits inside the same project: the `papers-with-notes` repository linked from the coverage table. That repository holds the complete accepted-paper lists as Markdown, without the notes. If you want to know whether a specific paper was accepted at all, that list answers the question faster than the notes do.

Update cadence and what maintaining a fork costs

The release notes show a steady rhythm through 2026: v1.0.0 in April with over 13,000 notes across nine venues, then ACL 2026, ICML 2026, CVPR 2026 and ICLR 2026 additions through June and July, ending with ECCV 2026 in v1.5.0 on 2026-07-11. The roadmap extends that forward, with EMNLP 2026, NeurIPS 2026, AAAI 2027, ICLR 2027 and CVPR 2027 listed by notification and conference dates. The last push recorded for the repository is 2026-09-06, which is consistent with an actively maintained project rather than a dormant one. For a consumer of the notes, upgrade cost is close to zero: pull the repository and you have the new content. For anyone maintaining a fork, the cost is the share-alike obligation plus the effort of re-merging a large Markdown tree on every content release, and the tree is large enough that conflicts are a real possibility. The cheap move is to consume the corpus as-is and keep your own annotations in a separate directory.

Editorial conclusion

Paper-Notes suits engineers and students who want a fast orientation pass over a conference before deciding which papers to read in full, and it suits anyone who wants the corpus as Markdown files under docs/ rather than as a website. It does not suit commercial use, because the README states the content is licensed CC BY-NC-SA 4.0, and it does not suit anyone who needs a verified summary rather than a reading aid. Before relying on it, open one paper page for a venue you know well, compare the note against the abstract, and check the papers-with-notes repository to confirm whether the paper you want is in the covered list at all.

Official sources

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