Large Language Models (LLMs) are becoming increasingly powerful tools for social science research. We present work in progress on in silico longitudinal surveys on LLMs with modular adapters.

This approach addresses issues of bias and prompt variation found in in silico surveys so far. While our initial implementation of this setup demonstrated abstraction capabilities of LLMs over digital trace data, validation on factual questions or seasonal trends is still required.

In the long term, in silico surveys could have the potential to enrich survey data gathered from humans by integrating abstractions over digital trace data or transcripts from interviews.

G. Ahnert, M. Pellert, D. Garcia, M. Strohmaier, Britain’s Mood, Entailed Weekly: In Silico Longitudinal Surveys with Fine-Tuned Large Language Models, Companion Publication of the 16th ACM Web Science Conference (2024) 47-50.