I don't think you even have to look deep into the model for it. There seem to be some projects who are simply analyzing the logits during decoding to estimate h

I don't think you even have to look deep into the model for it. There seem to be some projects who are simply analyzing the logits during decoding to estimate how "certain" the model is of the next token. But if all people do is random sampling or blindly taking the highest-scored prediction, this will of course fall under the table.