Getting arXiv Source Files: Why the TeX Matters for Reproducibility

An arXiv paper's PDF hides its structure. The source bundle exposes the preamble, the macros, the exact notation, and sometimes the comments the authors left in. For anyone reproducing a result or extending a proof, the TeX is more useful than the rendered page. Fetching it is a single link, but knowing what to look for in the bundle is the skill.

Where the source lives

On the abstract page, use the "TeX Source" link under Additional metadata, or construct the URL: https://arxiv.org/e-print/{id}. For older papers this may be a DVI or a gzipped tar; for newer ones a single .tex or a tar archive.

Download, extract, and open the main file named in \documentclass or in the submission metadata. Large collaborations often have dozens of files and an \input tree.

What the preamble tells you

The macro definitions are the paper's private language. A symbol that looks standard in the PDF may be redefined. Check \newcommand and \def lines for notation that changes meaning, and check \usepackage for packages that alter theorem formatting or citation behavior.

Custom theorem environments often hide assumptions: a \begin{assumption} block may state regularity conditions that the PDF renders as plain text.

Verifying claims against the source

Tables and figures in the PDF are rendered from source values. When a number matters, find the table in the .tex and check the raw input. Rounding differences and unit conversions sometimes appear between the source and the caption.

For proofs, the source reveals omitted steps compressed into \begin{proof} blocks, and comments like % TODO that signal incomplete arguments.

Figures, data, and the ethics of reuse

Source bundles occasionally include the plotting scripts or CSV data used for figures. Look for a data/ or figures/ directory. When present, that is the most reproducible artifact the paper offers.

If the bundle contains only PDF figures, do not assume the underlying data was released. Check the paper's data availability statement separately.

Reading the source is normal scholarly practice. Reusing substantial text or macros requires attribution, and the paper's license governs redistribution. arXiv licenses vary from permissive to all-rights-reserved.

A practical first check

Before building on a result, open the source, confirm the main file compiles with a modern TeX distribution if feasible, and note the version you used. A paper that fails to compile today is a signal that reproducing its experiments will require more than reading.

Working with the bundle, and reading macros as assumptions

Most bundles contain a main .tex file, a bibliography, and figures. Open 00README if present; older submissions sometimes include processing instructions that explain missing files.

Compile with pdflatex or lualatex depending on the document class. Missing packages are common; a modern TeX distribution resolves most, and a failed compile is usually a font or a class file rather than a corrupted source.

Diff the compiled PDF against the published version. Differences reveal the copyeditor's changes and, occasionally, content that changed after the arXiv version.

Custom notation often encodes constraints. A macro like \newcommand{\norm}[1]{\lVert #1 \rVert} looks standard, but a domain-specific variant may imply a particular space or norm.

Comments in the source frequently explain choices: % assume sigma-finite here, % reviewer 2 asked for this. Those comments are not part of the paper but they reveal the authors' own uncertainties.

Reproducing an experiment, and citation and licensing

Look for a data availability section and a repository link. If the bundle includes scripts and data, check the versions of the dependencies. Reproducibility failures are usually environment failures, not scientific ones.

If the code is absent, the method section is the only source. Compare it with the source's comments; authors sometimes describe an implementation detail in a comment that never made it into the prose.

Read the license on the abstract page. Some arXiv versions permit reuse with attribution; others reserve all rights. The license on the arXiv version may differ from the published version.

When citing, cite the version of record if it exists. Mention the arXiv identifier when the preprint differs materially.

Collaborations, versions, and the rest of the bundle

A browser extension that fetches source bundles should not auto-compile untrusted LaTeX. TeX can execute commands in some configurations, and a malicious bundle is a real risk. Compile in a sandbox or with shell escape disabled.

Open the source, locate the theorem or table you depend on, and confirm the statement matches the PDF. Then check the version history on arXiv. A result that changed between versions is a result you should cite carefully, with the version number.

A bundle with hundreds of files and a complex preamble can take time to understand. Start from the main file and follow \input and \include directives.

Check for a Makefile or build script; it documents the intended compilation order and engine.

Compare the arXiv version with the published version. Copyediting can change notation, and occasionally a result is corrected.

Cite the version of record when it exists; mention the arXiv identifier when the preprint differs.

Disable shell escape when compiling untrusted source. A malicious bundle can execute commands otherwise.

Run compilation in a container or a restricted user account for unknown sources.

The .bib file shows which version of a reference the authors used. That matters when a cite is disputed.

If the bundle includes scripts, check whether they run with the stated environment. A missing dependency is a reproducibility finding worth noting.

Not every submission includes LaTeX. Authors who use Word produce no source bundle, and older papers predate the requirement.

Fall back to the PDF and note the limitation in your reproduction attempt. Do not assume the absence means the authors withheld something.

The PDF from arXiv is built from the author source at submission time. If you recompile the source with a newer TeX distribution, the output can differ in page breaks and line numbers.

Cite the arXiv version number when quoting a page. Version two recompiles differently from version one, and a reader checking your citation will otherwise land on a different page.