We are almost certainly going to see lots of additional tokens added to vocabularies (like the thinking token, but also could be a "<LOGIC FAIL>" token), lots of sophisticated decoding strategies etc. Just need to generate the data.