Have you personally reviewed for big conferences or submitted and received reviews? It's a very noisy process that does toss out the lowest effort clueless stuff, but doesn't discriminate all that well between "meh" and "interesting", junior reviewers (the bulk) want proof of blood, sweat and tears. They want novel model modules and algo tweaks and complain about novelty that it's just A plus B, missing the point... They surely don't catch wrong results or incorrect claims because the catastrophic problems that invalidate papers are often in the implementation, not the nice math equations that motivate it.

In other words, Arxiv is what you use when you want to inform yourself on new research, conferences are for furthering your career by getting closer to your PhD graduation, expand your CV etc. And then to network and mingle with researchers in person and try to get hired.