You should probably think about why you think that making the model output reasoning steps which lead it to correctly answer questions it couldn't before is not somehow equivalent to reasoning.