Mathematics

LLM Survey Simulation (UQ)

Per-(LLM, survey-question) simulated human responses from "Uncertainty Quantification for LLM-Based Survey Simulations" (ICML 2025). Eight LLMs are prompted with random human personas to simulate responses on two real survey banks: EEDI (412 math multiple-choice questions, binary correct/incorrect) and OpinionQA (385 Pew opinion questions, ordinal sentiment score in {-1,-1/3,0,1/3,1}). Each model emits K simulated responses per question, one per synthetic persona, stored as separate trials.