This is driven by the idea that since LLMs are probabilistic engines during inference that, if the LLM "knows" or can "surmise" the correct answer, it will have a higher probability of being generated and thus self consistency methods can more reliably tease it out.