LLMs Achieve 52% Accuracy in Cross-Survey Human Respondent Prediction
July 7, 2026
Zero-shot LLMs (27B-120B parameters) can predict individual survey responses on unseen questions with 52% accuracy. This cross-survey transfer method shows LLMs can approximate supervised random forest models within 6 percentage points using existing respondent data.
HOW THIS AFFECTS YOU
●
researcherThis provides a more rigorous framework for evaluating LLM silicon sampling beyond simple distributional matching.