How to Submit to arXiv: A Practical Guide for Graduate Students

1. Why You Should Post to arXiv (and When Not To)

Posting on arXiv plants a public timestamp on your work, gets it cited

months before official publication, and makes your CV findable to hiring

committees who Google candidates the night before flyouts. In ML, physics,

and a growing list of fields it is now the de-facto first venue — many

papers accumulate more citations from the preprint than the published

version. The two situations where you pause: when your target venue runs

double-blind review and prohibits public preprints during the review

window (some HCI and medical venues), and when the result is patentable

and your tech-transfer office has not signed off. Check the venue policy

page before you click Submit, not after.

2. Getting Endorsed: The One Surprise for First-Time Submitters

arXiv requires an endorsement the first time you submit to a category

(e.g., cs.LG, stat.ML, physics.ins-det). Endorsement is not peer review —

it is a one-line vouch from someone already active in that category that

you are a legitimate researcher. Your advisor almost always qualifies; if

not, any senior co-author or recent collaborator does. The process: create

your arXiv account with your institutional email (this matters — `.edu`

and `.ac.kr`-style addresses bypass endorsement in many categories), start

a submission, and arXiv will tell you whether you need endorsement and

give you an endorsement code to forward. Allow 24 hours of slack —

endorsements are not instant.

3. Choosing Categories and Cross-Lists Correctly

Pick the primary category that best matches the paper's main contribution,

not the field you most identify with. A graph-learning paper aimed at

ML researchers goes to cs.LG primary, not cs.SI. Cross-list (secondary

categories) only when a substantial audience there would benefit — two

cross-lists is normal, four signals desperation. Look at five recent

papers you cite and copy their category pattern; the community already

decided where this work belongs. Choosing the wrong primary category does

not block your submission, but it does cut your day-one daily-listing

readership by half — most subscribers only read the primary list.

4. Preparing the Source: PDF-Only vs Full LaTeX

arXiv prefers LaTeX source, and posting source has two real benefits:

your paper gets HTML and ar5iv versions for accessibility, and a

"browse source" link other researchers actually use. PDF-only is allowed

but flagged. If you upload source, pack it like this: one flat tarball

(`tar czf paper.tar.gz *.tex *.bib *.bbl figures/`), a single main

`.tex` file (rename it to something obvious like `main.tex`), the `.bbl`

file alongside (arXiv runs latex, not biber/bibtex, so include the

compiled bibliography), and only the figure files you actually reference.

Strip out backup files, comments with reviewer responses, and anything

under a folder called `_old/`. The most common rejection: a non-portable

font or a custom `.cls` file the system does not know about — upload

the class file with your source if it is not on arXiv's allowlist.

5. Title, Abstract, and Metadata: The Part Google Indexes

The title and abstract you paste into arXiv's metadata fields are what

Google Scholar, Semantic Scholar, and arXiv's own search index — not

the version baked into the PDF. Treat the metadata as primary. Title:

same as your paper, plain ASCII if possible, no LaTeX commands beyond

simple math. Abstract: paste the published abstract; reflow paragraphs

so it is one block (arXiv collapses line breaks but not paragraph breaks).

Authors: full names, in the order you want cited. Comments field: this

is where you write "Accepted to NeurIPS 2026; 12 pages, 5 figures" or

"Extended version of our COLM 2025 paper". Use it — citation counters

and CV-stalkers read it.

6. License: Pick Once, Live With It

arXiv asks for a license at submit time, and you cannot downgrade later.

The four common choices: arXiv's non-exclusive license (default; lets

arXiv distribute but you retain copyright), CC BY (most permissive;

anyone can reuse with attribution — required by some funders), CC BY-NC-SA

(no commercial reuse), and CC0 (public domain). Default to CC BY unless

your funder, employer, or future-journal contract forbids it. The

arXiv-default license is fine but limits reuse in ways that bite later

(textbooks, tutorial papers, anyone wanting to mirror your figures).

Check your funder's open-access mandate — NIH, Wellcome, ERC, and most

European funders require CC BY for grant-funded work.

7. After Submission: Replacements (v2, v3...) and Withdrawals

Posting v1 is not the end. You replace your paper to fix typos, add the

camera-ready version after acceptance, or update results. Three rules:

do not replace daily (each version emails subscribers and looks chaotic;

batch fixes), always update the Comments field when you replace (`v2:

camera-ready NeurIPS 2026 version, fixed Theorem 3.2 proof`), and never

replace to remove an author silently — that gets noticed. Withdrawals

are different: you can withdraw a paper, but the version history stays

public forever. The right reasons to withdraw: a fatal error you cannot

fix, or the work has been substantially absorbed into a different paper.

The wrong reasons: a journal rejected you, you found a typo, your advisor

changed their mind. arXiv is an archive — treat it as such.