They seem, in theory, to work quite well. In reality, though, the 'adaptive learning' algorithms are simply "answer x correctly, go to y; answer x incorrectly, go to w" sort of things. They have fancy charts and graphics showing 'knowledge clusters' and 'though pattern analyses', but really, it boils down to my original statement. I've tested, re-tested and counter-tested these programs using various skill levels and completely random answers.
X is right, proceed to y. X is wrong, proceed to w.
From the 2-3000 students I've seen run through these courses in the last four years, there are three types of interaction with this model: those that love it and do well (but would also love and do well in a structured classroom), those that hate it and do poorly (but would also hate and do poorly in a structured classroom), and those that have absolutely no idea why they're being forced to work with a computer that doesn't allow you to answer questions.
I may be a bit old fashioned in my pedagogy and educational theory (yet I'm only in my 20's), but there really do seem to be students that simply respond better to being able to ask questions and receive answers from a living-breathing human being.
Really, the only benefit outside of cost savings to employees and administration of universities and colleges is that these programs allow us to find the exact point at which a particular student's understanding begins to break down. That's new, and fascinating (and brings up whole new concerns on privacy in my opinion).