The closest explanation to how chain of through works is suppressing the probability of a termination token. People have found that even letting llms generate

The closest explanation to how chain of through works is suppressing the probability of a termination token. People have found that even letting llms generate gibberish tokens produces better final outputs. Which isn't a surprise when you realise that the only way a llm can do computation is by outputting tokens.