Got interested and tested on my object Sox2 (binds DNA with a HMG-box domain). HMG-box might bind RNA, but not as its primary function (so strong coevolutionary signal is unlikely), and as far as I know no HMG-box-RNA complex exists in PDB. But AF3 still predicted pretty similar structures

I'm no expert, but can try to answer. While according to the AF3 Supplementary all standard residues are represented as distinct tokens, AF3 also employs atomic features (derived from reference conformers) in addition to token features. So it should be able to recognize chemical similarities