(1) the thumbnails indicate that the PDF is available and can be downloaded. This is to promote submissions from pre-print servers and open access journals that make their research readily available. If the PDF is behind a paywall there will not be a thumbnail.
(2) the behavior of the site differs slightly from HN and Reddit. Clicking the title link of a publication takes you to be comment page, whereas the thumbnail takes you to the PDF. Open to hearing what people think about this, but the thumbnail image replaces the title link.
(3) the thumbnails do tend to look similar - but you can often tell what journal it is in and you can recognize familiar papers from it. I like it when authors put figures up front too which some of the ML paper do and which adds some variety.