How to Submit a Paper to arXiv: LaTeX Submission Guide

The short version: upload your LaTeX source (not just the PDF), make sure it compiles on a machine that is not yours, include the .bbl file, and hit submit before 14:00 Eastern if you care about when the paper announces. Everything below is the detail that separates a five-minute submission from an evening of cryptic compile errors.

Before anything: your account and endorsement

arXiv is not a free-for-all. The first time you submit to a category, you may need an endorsement from an established author in that category. In practice most people never see this step: registering with an institutional email from a recognized university or lab usually qualifies you automatically for the categories your institution publishes in.

If you do get asked for an endorser, the submission page gives you a code to send to someone who has published in the target category. Your advisor or a co-author is the obvious choice. It takes them two minutes. Do this a few days before your deadline, not the night of; endorsers are humans with inboxes.

What you actually upload

If your paper is written in LaTeX, arXiv requires the source, not a compiled PDF. This is policy, not a suggestion: the system detects PDFs produced by TeX and rejects them. arXiv compiles your source on its own servers with TeX Live, so the output readers see is what arXiv built, not what your laptop built.

Upload a zip (or tar.gz) containing:

• the main .tex file and any files it \inputs or \includes
• every figure file the document references, in the exact paths it references them
• the .bbl file (more on this next)
• any custom .sty or .cls files that are not part of standard TeX Live

Leave out the junk: .aux, .log, .synctex.gz, editor backups, the 40MB of figures you cropped out three drafts ago. arXiv processes what you give it, and stray files cause more failures than missing ones.

The .bbl rule

This is the classic first-submission failure. For most of arXiv's history the rule was blunt: arXiv does not run bibtex, so if you upload a .bib file your citations all come out as [?]. The modern pipeline is slightly smarter (if you include your .bib files it can now run the bibliography step itself), but the reliable path is still the old one:

1. compile locally so bibtex generates main.bbl
2. upload that .bbl alongside main.tex

When a .bbl is present, arXiv uses it verbatim: you control exactly what the reference list says, and nothing depends on which bibliography tool the server picks. One catch on the newer pipeline: the .bbl should come from a reasonably current TeX distribution, since arXiv's TeX Live checks the format version of the file. If your local install is years old, update it before generating the final .bbl. The third option, pasting the thebibliography environment directly into your .tex, also works and never surprises anyone.

The compile failures everyone hits

Missing or misnamed figure files

The number one failure. arXiv's servers are case-sensitive: Figure1.PNG and figure1.png are different files there, even though Windows and macOS will happily pretend they are the same on your machine. If the compile log says a figure is missing and you can see it in your zip, check the capitalization first.

Absolute paths

\includegraphics{/Users/you/paper/figs/plot.pdf} compiles on exactly one computer in the world. Every path in your source must be relative to the main file: \includegraphics{figs/plot.pdf}. The same goes for \input paths.

Shell-escape is disabled

arXiv compiles without -shell-escape, for obvious reasons: it is arbitrary code execution on someone else's server. Anything that depends on it breaks, and the most common casualty is minted for syntax-highlighted code. Swap it for listings, or pre-generate the highlighted output locally and include the result. Externalized TikZ (\tikzexternalize) fails for the same reason; pre-compile those figures to PDF and include them as graphics.

The engine-detection edge case

arXiv inspects your source to decide how to compile it. It almost always guesses right, but there is one edge case worth knowing: a paper that mixes signals (say, an old EPS figure left over next to PNGs) can get routed down the DVI path, where your PNG and PDF figures fail. Older guides tell you to force the issue with \pdfoutput=1 in the first lines; arXiv's current documentation says not to override the output format, and the honest fix is to remove the mixed signals: use one figure format family and let the detection work.

Figures

PDF, PNG, and JPG are all fine for a pdflatex paper, which is nearly every paper today. EPS belongs to the legacy latex→dvips route, and arXiv does not convert figures between formats during compilation. If your figures are EPS and the rest of your setup is modern, convert them to PDF yourself before uploading (any vector tool or a local epstopdf pass does it) rather than betting on the legacy path.

Keep figure sizes sane. There is a limit on total submission size, and a 30MB rasterized plot helps no one. Export vector PDFs for line plots, compressed PNGs for raster images.

Ancillary files

Code, data, videos, or long supplementary material that is not part of the compiled paper goes in a directory named anc/ at the root of your upload. arXiv lists these as downloadable “ancillary files” alongside the paper without trying to compile them. It is the right place for a README and a snapshot of your experiment code, better than a link that dies when you change universities.

Choosing categories

You pick one primary category (say, cs.LG) and optionally cross-lists (say, stat.ML). The primary determines which listing your paper leads in and which moderators review it; cross-lists get you into the daily mailings of adjacent communities. Two or three well-chosen categories is normal. Six is a flag: moderators do notice cross-list stuffing, and it can slow your paper down.

The announcement schedule (and why people time submissions)

arXiv runs on Eastern US time. The cutoff is 14:00 ET on weekdays. Submit before the cutoff and your paper is announced at 20:00 ET the same weekday; miss it and you roll into the next announcement. There are no announcements Friday or Saturday night; anything submitted after Thursday 14:00 waits until Sunday or Monday evening.

Why do people care? Position. Papers submitted right after a cutoff land near the top of the next day's listing, and the top of the listing gets more eyes. There is real (if modest) evidence that listing position correlates with early citations and downloads. If you have ever wondered why a flood of ML papers appears at 14:00:05 ET, that is why. It is a small edge; a good abstract is a bigger one.

Licenses: read this once, choose deliberately

At submission you grant arXiv a license to distribute the paper. This is perpetual and irrevocable, which is the whole point of arXiv as a permanent record. The choice is which one:

• arXiv's minimal license: the default and what most people pick. arXiv can distribute the paper; you keep all other rights. Journals are used to it.
• CC BY 4.0: anyone can reuse with attribution. Pick this if your funder mandates open licensing, and only after checking your target journal accepts submissions already under CC BY.
• CC0 and the other CC variants exist, but they are rarely the right call for a paper headed to a journal.

When unsure, take the default. You can apply a more permissive license to a later version, but you can never un-apply one.

Versions, replacements, and the myth of deletion

Found a typo an hour after announcement? Relax. Replacements are cheap and completely normal: submit a v2 and it announces on the next cycle with a “replaced” note. Serious papers routinely reach v3 or v4 as reviews come in. Nobody counts your versions against you.

Moderation holds: don't panic

Occasionally a submission goes “on hold” instead of announcing. This means a moderator wants a look, commonly because the category choice looks off, the paper resembles an earlier submission, or something tripped an automated check. It is not a rejection and usually not about your paper's quality. Most holds clear in a few days to a couple of weeks with no action from you. Don't email support on day one, and never create a second account to resubmit. That genuinely will get you in trouble. If a conference deadline depends on the arXiv timestamp, submit early enough that a hold cannot sink you.

Pre-submission checklist

1. Endorsement sorted (or institutional email registered) days before you need it
2. Fresh local compile from a clean directory: delete the .aux files and build twice
3. .bbl generated and included; no reliance on bibtex running server-side
4. All figure paths relative; filenames match on case exactly
5. No minted, no -shell-escape dependencies, no externalized TikZ
6. Figures in PDF/PNG/JPG; EPS converted beforehand
7. Supplementary code and data in anc/
8. Primary category chosen; cross-lists kept honest
9. License picked deliberately (default is fine)
10. Submitted before 14:00 ET, with margin for a moderation hold if a deadline depends on it

Two related guides: how to cite arXiv papers covers the reference side, and if you are starting from a PDF that was never LaTeX in the first place, our PDF to LaTeX converter gets you compilable source to build on.