The models struggle with reasoning ... period.

Any apparent reasoning ability is really a statistical simulation thereof drawing on similarity with training data.