Tag
This paper proposes using fully synthetic LLM-generated prompts to improve ecological validity in measuring LLM political stance, extending IssueBench beyond writing assistance to information seeking and opinion sharing tasks, and showing that templated prompts may overstate models' leanings.