PDFs dominate AI research paper distribution

PDFs still rule how AI papers move from author to reader. Most of the time, the paper that people download, skim, forward, and cite is a PDF, even when the source began as LaTeX, TeX, or another editable file.

That sounds plain, but it matters. A PDF is stable. It keeps the layout, figures, equations, and page breaks in place. For a field that lives on tables, formulas, and dense charts, that stability is not a small thing. It is the reason the same paper can show up cleanly on a laptop in Boston, a phone in Bangalore, or a printout on a desk.

arXiv makes this pattern easy to see. For most articles, PDF is the top download link. The service also says it generates PDF for TeX and PostScript submissions, and that papers submitted in PDF are available only in PDF. In practice, that means the PDF is not just one option among many. It is the main public face of the paper.

That is the key fact behind the headline. AI research papers are often written in source formats, but they are usually shared and consumed in PDF form. The source may live behind the scenes. The reader usually meets the paper as a finished document.

This is not only an arXiv habit. Publisher portals and open repositories also tend to present papers first as PDFs. That has been the default academic distribution pattern for years because it works across systems and preserves the author’s layout. In AI, where preprints move fast and conference timelines are tight, that default becomes even more visible.

There is also a human reason PDF survives. It fits the way research is judged. Reviewers want the same page view. Engineers want to copy the method section without losing structure. Managers want the charts to look the same when a link is opened in a browser, then again in a meeting deck. The format reduces friction, and in a deadline field, friction loses.

Still, the format has a cost. PDF is strong at preserving a page. It is weaker at being a living, web-native document. Accessible reading can be harder, especially for screen readers. Search, reflow, and machine parsing can also be awkward when a paper is treated like a scanned page instead of structured text. arXiv itself has acknowledged that the PDF-first model helps distribution while creating accessibility limits.

That is the part worth sitting with. The format that makes papers easy to move also makes them harder to change. It serves speed, stability, and citation. It does not serve every reader equally well. The gap is not theoretical. It shows up in accessibility, text extraction, and the basic fact that a paper built for a page is not the same thing as a paper built for the web.

I think this is why the PDF has stayed so stubborn. Research communities like formats that feel final. A PDF looks finished. It looks safe to share. It looks like the version that counts.

But the final look can hide the real workflow. Many AI papers begin as code-like source files, then get compiled into PDF for public release. That split matters. The source may be where the work is made. The PDF is where the work is distributed. For most readers, that is all they ever see.

The practical consequence is simple. If someone asks what format AI research papers are usually in, the answer is PDF. If someone asks why, the answer is that it remains the most portable and familiar way to distribute a paper without breaking the layout that makes the paper usable. The format is old, but the system around it still depends on it.

I do not think that is an accident. It is a compromise that has lasted because it serves publishers, repositories, reviewers, and readers at once. The unresolved part is whether that convenience should keep outranking accessibility and web-native formats. That question is still open, and it will stay open as long as research culture keeps valuing a clean page over a more open document.

After the Demo is where that tradeoff shows up most clearly, in the gap between what looks polished and what actually works for the people who have to read, share, and live with it.

Browse: AI Research and Benchmarks