One way that it works for programming is doing small exercises and then checking the answers. For instance, the 99 Problems in Prolog (or its translations to Lisp, Haskell, etc):
http://www.ic.unicamp.br/~meidanis/courses/mc336/2006s2/func...
But as you say, for larger problems it's harder to find good "answers". Even for something relatively small like unix utilities. Say you want to write cat or echo. You get source code from BSD and GNU and Solaris and they're quite different from each other (I recall seeing a comparison somewhere, mainly putting the GNU code in bad light, perhaps unfairly. Anyone has the link?).
Really polished code is hard to find. I've been playing around off and on with Ken Thompson's regular expression search paper most recently.
If this is true, could studying (reading) good code do the same for programming? One could, for example, begin to write an application and find code to a similar application where they would be able to check why things are implemented a the way they are as they go along. The tricky part here is deciding which code is worth studying.