Reproduce “Specializing Large Language Models to Simulate Survey Response Distributions for Global Populations” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yong Cao, Haijiang Liu, Arnav Arora, Isabelle Augenstein, Paul Röttger, Daniel Hershcovich

The authors fine-tune language models to predict population-level answer distributions for a country and survey question rather than individual responses. Using option probabilities and KL loss, they improve matching on World Values Survey data, including held-out countries and questions. For Llama3-8B-Instruct, average 1−JSD rises from 0.613 to 0.823, but unseen questions remain harder and predictions less diverse than human responses.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.