Abstract

The emergence of large language models as in-silico subjects for social science poses a central question: can they genuinely simulate diverse human behavior, or do they merely produce plausible, homogenized artifacts? This talk demonstrates that LLMs are powerful but imperfect simulators by presenting three core contributions. First, we establish the “Persona Effect,” showing that persona-prompting a 70B model captures 81% of explainable variance in subjective tasks, creating a strong baseline for individual-level simulation. Second, to address data scarcity, we introduce iNews, a large-scale dataset of personalized affective responses to news, enriched with persona information. Finally, we introduce SimBench, the first large-scale benchmark for group-level simulation, which reveals the strengths and critical weaknesses of current models. I conclude by arguing for the specialized datasets and training required to advance the frontier of high-fidelity human simulation.