I think sites like hacker news and slashdot have a pretty good handle on this problem but only in their own communities. A good start to a more generalized solution would be to make these currently siloed reputation systems portable.
Although maybe you could train algorithms to detect derogatory nouns and adjectives. And you could still allow reps on individual comments and hide low comments by default, which effectively crowdsources the troll moderation. I believe I've seen sites using Disqus in a way close to this.
The time it takes to create a login is trivial compared to that necessary for a decently detailed comment.
To borrow a concept, it's a bit like Whuffie in Doctrow's Down and Out in the Magic Kingdom.
Even look at Reddit - trolls are usually downvoted, sure, but a lot of people with high reputation are really just novelty accounts. That's not bad for Reddit, but that's not what I would associate with high reputation on HN.
And finally, look at one of the comments below about gaming the system - that's the problem with having a (displayed) numeric value associated with one's reputation.
So someone making an honest attempt to discuss unpopular opinions gets shut out but someone making relatively meaningless platitudes to the group gets cheered on.
For a small group that isn't interested in those opinions anyway it wouldn't matter, but it limits the usefulness of that reputation elsewhere.
That's the Achilles heel of reputation moderating systems. You don't have to be insightful; you just have to play to the crowd.