You just need better training data. Your internal language model can't follow a non-linear argument or understand basic logic.