You could easily fix this by weighting the value of each new vote based on the position the comment was in when it got the vote. These weightings would need to be calibrated but would produce a better outcome.
I think it would also have to be logarithmic, because being the top comment is MUCH better than being the third comment, while being the 6th comment is only marginally better than being the 8th comment.
Coupled with the stochastic model, it could make it a far more level game.
I also think it's easier to calibrate [an algorithm that estimates which comments people have read], than to calibrate [how much more importance to give to a vote that happens far away from the original post].
In fact, the latter is impossible? Because you have nothing to calibrate against? You don't know what the correct result is.
But here's a reasonable (?) calibration for an algorithm that estimates which comments people have read: If you upvote comment X, then you have probably read its 3 earlier siblings and its parent and grandparent. This doesn't need much calibration, and would probably (I think) work fairly well on the whole? — It's somewhat possible to validate and tweak this calibration, by observing how people actually do, when they read forum & blog comments.