Unsupervised decoding of encoded reasoning using language model interpretability