My dream would be arxiv hosting a reddit-style upvote/downvote discussion thread for every paper. I've seen this work beautifully on some science subreddits, as well as on Scot Aaranson and Terence Tao's blogs. It can really help flesh out problems/questions, generate intelligent discussion, and even lead to cooperation between participants on further research. One requirement for it to work is strong moderation to keep the discussion focused.