Simple answer: the peer-review process can be slow and not all papers deserve to be accepted at (major) conferences and journals (e.g. some are just tutorials or simple or more complete descriptions of previously published ideas, so they do not propose anything novel, which is required for being accepted at major conferences or journals).
Why is peer review slow? Because multiple people need to read the paper carefully, understand it (so they may need to read about the topic, if they are not fully familiar with it), analyze it, and maybe read (previously published) related literature to know if the proposals are really novel or deserve to be published (if you're proposing a VAE, you will not go anywhere, as VAEs have already been proposed and accepted at ICLR 2014, i.e. no progress and novelty!).
I found several papers on arXiv to be useful, but I also found low-quality ones. Generally, if you can cite a paper that has been accepted at a major conference or journal in your field, you should do that, rather than citing a pre-print. If a paper has been peer-reviewed, then it's been scrutinized in multiple senses (although some still contain some typos), so it's very unlikely to find a paper that has been accepted, for example, at NeurIPS that is poor (e.g. the GAN or transformer papers were accepted there). You will find so many useful papers published there or in conferences/journals like ICLR, JMLR, and so on. You should also notice that some papers that are first published on arXiv might be later presented/accepted at some conference or journal (example: VAE), so the citation may redirect you to the pre-print version, but this does not mean that there isn't a version that was later published in some journal. People may first publish on arXiv so that other researchers can have early access to the ideas, and so progress can be faster, but, again, there are many papers on arXiv that will never be accepted at any (major) conference/journal.
(By the way, although this question was asked in the context of deep learning (hence Artificial Intelligence), so it can be considered on-topic here, you probably would find more useful answers at Academica Stack Exchange, because this topic is not unique to research in AI (e.g. there's also bioRxiv, which has a similar role to arXiv, i.e. proliferation of ideas, etc., but for biology). In fact, you can already find some answers there that address this question, for example, this one.)