“Users on the LM Arena platform typically spend only a few seconds reviewing model responses before voting, leading them to prefer superficial qualities over accuracy or instruction-following.”