Wired Magazine reports that on the Caltech-MIT Voting project(you can follow the Wired link to a pdf file) that the 2004 electoral returns showed an improvement on 2000 with respect to ballots scored as "residual," that is, without identified a presidential preference. Estimates run that about 1% of voters intentionally do not vote for president, but returns above that percentage suggest counting problems. The measure for residual votes were 1.1% in 2004, compared to 1.9% in 2000.
Wired reports that credit is due to, believe it or not, throwing money at the problem. States which improved ballot design and implemented tech upgrades (including Florida) showed the best drop in residuals.
There's still a lot to do, especially right at the polling place. Check out this older Tech Review piece by Ted Selker of the Voting Project for examples of poll place sloppiness.