lmarena user prompts are sent to two different randomly selected models and the results are shown to the user, and they vote for whether they prefer model A or model B. With many votes from many users, user preferences can be derived. From the lmarena page linked above