I consider it a heresy that some sites (fuck ResearchGate, Academia.edu) consider it fine to alter the PDFs. Document integrity should be the holy grail. Files should be final and never get touched. If they are, they are different.
Different versions of a paper should have different DOIs. A DOI is fundamentally quite abstract, and there is no reason that an object-appropriate versioning scheme can't be laid on top. Figshare, for example, uses a versioning system where there is a single DOI referring to the overall artefact (e.g. 10.6084/m9.figshare.2066037) and a versioned-by-suffix DOI referring to a specific version (e.g. 10.6084/m9.figshare.2066037.v16). That works really well IMO.
arXiv:1707.08475 [stat.ML] (or arXiv:1707.08475v1 [stat.ML] for this version)
http://sci-hub.cc/10.1021/acs.jpcc.7b06335
It's already passed peer review, but the final version will be tidied up with additional editing and will have e.g. the line-by-line numbering stripped out. I see sci-hub versions of papers from months ago, or older, that never update to the final edition. The content doesn't differ much but the final version may have better-placed figures, corrected typos, and other minor improvements.