Reading a paper

arxiv paper takes any reference arXiv would recognise.

arxiv paper 1706.03762
arxiv paper arXiv:1706.03762v7
arxiv paper hep-th/9711200
arxiv paper https://arxiv.org/abs/1706.03762

To see how a reference was understood without spending a request, use arxiv id.

arxiv id "arXiv:1706.03762v7"

That parses the id, the version and the style locally and asks arXiv nothing.

Depth by depth

The default is --depth meta, which is two requests on the api plane.

arxiv paper 1706.03762 --depth quick   # the export API alone
arxiv paper 1706.03762 --depth meta    # adds the report number, MSC and ACM classes, licence
arxiv paper 1706.03762 --depth full    # adds versions, submitter, category names, capabilities
arxiv paper 1706.03762 --depth text    # adds affiliations, the section tree, the bibliography

full and text cross onto arxiv.org at fifteen seconds a request, so they are a different order of cost. See depth for the numbers.

What a read did not look at is on the record.

arxiv paper 1706.03762 -o json | jq '.[0].missed'

The version history

A paper at v7 and a paper at v1 look identical on the export API, because it publishes no version history at all.

arxiv paper 1706.03762 --depth full -o json | jq '.[0].versions'

Each version carries its date, its size and its source type, out of arXivRaw and the abstract page together.

Pinning a version in the reference reads that version.

arxiv paper 1706.03762v1

Withdrawn papers

A paper the author withdrew is still on arXiv, with its id, its title and its history. What changes is that a short new version goes on top and arXiv marks that version withdrawn.

$ arxiv paper 0902.4054 --depth full -o json | jq '.[0] | {withdrawn, via: .via.withdrawn, comment}'
{
  "withdrawn": true,
  "via": "s3",
  "comment": "This paper has been withdrawn"
}

withdrawn is read from the marker in the submission history and not from the comment, because comments are prose. They run from Withdrawn to four sentences of explanation, and v1's main result and mechanism claim are withdrawn is a comment on a paper that is still very much there.

Each version says whether it is the withdrawn one, so a paper withdrawn at v2 and replaced at v3 reads as a paper that is there.

It costs --depth full, because only the abstract page carries the marker. There is a second route, an OAI header saying deleted, which is arXiv removing a record rather than an author withdrawing a paper, and it comes back as "via": "s2" when it happens. It is rare enough that 40,000 OAI headers sampled on 2026-08-15 contained none.

Files

arxiv files 1706.03762

Which files exist is already on the paper record as two capability flags off the abstract page, so the list itself costs nothing beyond the paper read.

A PDF is always there. HTML is there for papers submitted since December 2023 and for some earlier ones. Source is there unless the paper was submitted as a PDF.

The size on a source file is arXiv's own figure out of the submission history, and it is labelled as such because it is not the number of bytes that will arrive. The version table says 1,102 KB for 1706.03762v7 and the archive served is 1,150,988 bytes. There is no figure at all for the PDF anywhere in the metadata.

arxiv files 1706.03762 --measure

--measure asks arxiv.org about each file and fills in the real size, the filename arXiv would give it, the content type and arXiv's sha256. It is one request per file on the fifteen second plane, and it is a one byte range request rather than a HEAD, because a HEAD comes back through the CDN with no content-length on it.

Downloading

arxiv download 1706.03762                        # the PDF
arxiv download 1706.03762 --kind source --extract

A file that is already there is not fetched again, and --force says to fetch it anyway. A download in progress is a .part file, which is what tells a truncated PDF apart from a whole one, and --resume continues it with a range request.

--extract unpacks a source archive into a directory named after it. An entry that would write outside that directory is refused rather than skipped, and a symlink is skipped because nothing in a LaTeX source needs one.

robots.txt disallows /src and /e-print, and arXiv's data policy says full content harvesting is not permitted except by arrangement. So this is one paper at a time, at the fifteen second pace, for a paper somebody named. There is no --all and there is not going to be one.

Full text

arxiv fulltext 1706.03762

arXiv renders LaTeX submissions to HTML, and that rendering is the only place arXiv publishes author affiliations, a section tree, or the body of a paper at all.

The read is the abstract page and then the rendering, because has_html is on the abstract page and it is the only honest way to know whether a rendering exists. That is two requests on the fifteen second plane, so the command takes half a minute the first time and nothing at all the second: a rendering never changes, so it is cached for a month.

A paper arXiv never rendered exits 7 and says so.

Maths comes back as the LaTeX the author wrote, out of the alttext attribute, because a downstream reader can parse that and cannot parse rendered MathML.

What the paper claims

arxiv edges 1706.03762
arxiv edges 1706.03762 --predicate authored
arxiv edges 1706.03762 --depth text --predicate cites

Each row is a subject, a predicate, an object and the surface behind it. See the claim graph.