i agree that it seems easier to simulate working conditions with a homework assignment. (it does add a lot of work for the company to do homework problems well, but let's leave that aside)
the downside is that the candidate has a limited amount of time to talk to you. they often have a full-time job going on top of the job hunt. secondly, are you really game to spend 0.5-2 days of unpaid work just for the chance of being invited to interview further?
An additional perspective I didn't think of until now is whether this type of recruitment process would be more beneficial to the industry as a whole. An argument could be made that we'd all be better off as a whole if interviews tended to more accurately assess candidates and candidates were able to get a more realistic sense of the type of work they'd be doing for a company.
I find this point interesting, as to me, this would be an argument for, not against. Time constraints on homework assignments are a lot closer to the constraints you'd have programming for a job than "whiteboard coding" would be. I'd hazard a guess that doing a good job on an in-depth, open-ended problem given a few days to a week is a better indicator of ability to be a successful software engineer than being able to solve a logic puzzle on a whiteboard in fifteen minutes.