Can LLMs Replace Survey Respondents?

Large language models can accurately replicate the average responses from economic surveys but fail to capture the diversity of human opinion, a problem known as mode collapse. Researchers tested 'unlearning' methods to make models forget memorized statistical data, which successfully restored a more realistic distribution of answers. One unlearning technique, Gradient Ascent, allowed a Llama-3 model to better simulate the results of a randomized controlled trial involving inflation expectations. While promising, the results show that unlearning is not a universal fix and that LLMs still struggle to fully imitate diverse human personas in survey simulations.