Comment by ydj

How do you determine that the model was reasoning in Chinese in layer X? I would think the middle layers do not map into any tokens.