LMSYS Arena blog explores factuality evaluation challenges

The post highlights that human preference rankings miss factuality, which is hard to evaluate manually. It hints at a new automated approach for fact-checking model responses at scale.