I think that as long as the attention mechanism has been trained on each possible numerical token enough, this is true. But if a particular token is underrepres

I think that as long as the attention mechanism has been trained on each possible numerical token enough, this is true. But if a particular token is underrepresented, it could potentially cause inaccuracies.