Paper on adaptive political surveys with GPT-4 published in PLoS One
My first paper as a PhD student! Very happy to have collaborated with Fynn to get this published in PLoS One.
We looked at how well GPT-4 can mimic user responses in a political survey, and how such synthetic data helps to train statistical models in adaptive questionnaires. This addresses the cold start problem, one of the main limitations of an adaptive questionnaire. We applied the method to the Swiss Voting Advice Application Smartvote, and showed that it can significantly increase the recommendation accuracy for early users.