Instead, they respond to some features in the simplify caricatures. The most likely of these features are mirror symmetry and optimal level of information density (amount of details in the picture), but there may be other, like smooth curves compared to sharp corners etc. In the experiemnts that show infants prefer faces to other pictures, the control pictures are always worse in some of these features. In many cases it is the mirror symmetry, but in other cases it is wrong information density, or unequal distribution of information density.
However, this phenomenon has a very simple explanation. The name of a person has very low correlation with any other of his/her attributes. Most significanly, all the other attributes of a person put together still gives almost no indication of what his name is. In contrast, other attributes tend to correlate with each other, at least weakly. Thus recalling ay other specific attribute is helped by recalling relating attributes, but recalling the name is not.
Some of the researchers in the area do incorporate this notion in their models. However, these models already assume that there is a specific mechanism for recalling names and other personal attributes, and the special difficulty in recalling names is implicitly used to support this assumption. As explained in the previous paragraph, there is no need for special mechanism. The only required assumption is that recall of a piece of information is aided by recalling related pieces of information, and this assumption is obviously true.