Perhaps it is getting a bit ridiculous publishing 900+ pages of a math proof in a pdf format. Perhaps mathematicians should move to a GitHub style publication platform.
I use a screen reader, and unless the pdf is explicitly created as an accessible pdf, it is just garbled up nonsense.
In this particular case <https://arxiv.org/format/2205.14808> lets you "Download source", where the source in question is a tar file with four png images and a large .tex file. The .tex file has long swathes of English text; equations are mostly in $math$, although there are some parts that are likely to be less straightforward to access (e.g. under "Definition of the Ricci and curvature coefficients").
arXiv admins and most authors will react supportively to contact from blind and visually impaired people who want to read a preprint.
Some categories and broader disciplines have more work to do than others. For example <https://astrobites.org/2022/07/17/making-astronomy-accessibl...> I think should link to the abstract page <https://arxiv.org/abs/2206.13815v1> so that one can find the .tex file (and a .docx), but instead directly links to the PDF.
I use a screen reader, and unless the pdf is explicitly created as an accessible pdf, it is just garbled up nonsense.
There's room for improvement over the status quo, but with the standards we have available today, faithfully representing the content requires a faithful representation of the visual layout. So a typesetting format like PDF is what's needed.
S-expressions?
LaTeX?
You expect to see
$a^2+b^2=c^2$
but you actually see manually spacing, sub/super-index before and after the object, weird hats, ...
$a+{}^2\!\hat{x}\ \leq& \ 7\ y_2\!\!^5+\epsilon$
https://arxiv.org/format/2205.14808
Huh. this account apparently got shadowbanned. Guess it's another good reminder not to post here.
For mechanical verification, PDF seems like the worst possible input format.
It's a flaw with latex as far as I'm concerned.