Silicon Sampling via Cross-Survey Transfer

This paper proposes cross-survey transfer as a stricter test of silicon sampling. An LLM receives a respondent’s answers to one set of questions and predicts answers to entirely unseen questions from the same survey, testing respondent-level coherence rather than only matching aggregate distributions. Using Taiwan Election and Democratization Study 2024 data, three open-weight models ranging from 27B to 120B parameters, and supervised baselines, zero-shot LLMs reached 52% accuracy on unseen items, within 6 percentage points of a same-population random forest. Predictability varied sharply by construct, from 67% for partisan attitudes to 23% for sovereignty.