If you factor things out into sub-functions that have no semantic meaning on their own then all you are doing is making the code harder to understand. You're also making it harder to maintain because of all the extra state that will need to be passed around between the sub-functions, which may change later.
For large complex functions that cannot be broken up sensibly, the paragraph splitting method is exactly what I use. Such large functions are a code smell and you should think carefully about whether it really does need to be so large, but there are indeed cases where it is the best option. Nobody should get too attached to dogmatic rules like "no function may exceed 25 lines".
The style that I prefer is slightly different to the GP though. I usually put about a paragraph of comments at the top explaining why the function is so long, giving an overview of the algorithm and other information that is inappropriate for a JavaDoc-style comment (since those are for the function consumer rather than a maintainer). I then give each "paragraph" of code a section heading and sometimes number these (especially if I have written out the algorithm in line-numbered pseudocode in the explanatory comment).
