Improving Self Consistency in LLMs through Probabilistic Tokenization