This was a very far-sighted move by, I believe, Paul Ginsparg back in the early days. It significantly increases the headache of submitting to the arXiv because you have to get your tex file to compile with the tex distribution on their servers rather than just on your own home box. But it makes the arxiv vastly more future proof than it would be if you could just upload PDFs.

I think arXiv / LANL preprint service predate PDF and the web (ftp and gopher). tex is seems more popular still in physics than other disciplines.

Yes, the arXiv started with PDF-less tex. (I believe people just compiled to postscripts files, which could be printed directly.) But when PDFs appeared, and especially as latex distributions became less standardized, there would have been a natural pressure to accept PDFs.

yes, this is why they have accepted PDF for some time https://arxiv.org/help/submit_pdf

(maybe for as long as since 2004 -- but ironically archive.org could not access arXiv for some years https://web.archive.org/web/20041101000000*/https://arxiv.or... )

Most other preprint services, such as those based on eprints (and in other disciplines) have always accepted PDF.

The reason they accept PDF is not because of the difficulty in getting tex files to compile on different distributions that I mentioned. Indeed, as they say on the page you link to, PDF files created from a tex file are specifically rejected by the arXiv.

Rather, PDF files are accepted for the (fairly small minority) of papers that are written using alternative editors like MS Word.

For any arXiv submission that was submitted via LaTeX, you can download the source.