LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals
The paper presents a large‑scale empirical study (N=1,052 U.S. participants) that builds generative agents using GPT‑4o and grounds them in participants’ self‑reports. Three grounding modalities are compared: (1) two‑hour semi‑structured interview transcripts, (2) structured survey data (General Social Survey, BFI‑44), and (3) a combination of both. Agents are evaluated on three downstream tasks – replication of General Social Survey (GSS) items, prediction of Big Five personality scores, and performance in incentivized economic games – and benchmarked against each participant’s two‑week test‑retest consistency. Regression analyses show modest demographic effects on predictive performance, with political ideology, party affiliation, and sexual orientation showing the largest disparities. Demographic Parity Difference (DPD) tables reveal that interview‑grounded agents achieve parity scores close to participants’ own consistency (≈82‑86% of test‑retest), outperforming demographic‑only or persona‑based baselines. Robustness analyses demonstrate that agents retain high normalized accuracy (≈0.82) even when large portions of interview data are removed, and that maximal agents (integrating interview, survey, and experimental data) achieve the best performance (normalized accuracy ≈0.87). The study also details the interview script (≈2‑hour, covering life history, health, finances, politics, and values) used to collect rich narrative data, and provides extensive demographic quota information to ensure a stratified sample. Overall, the findings suggest that grounding LLM agents in rich self‑report data yields scalable, general‑purpose simulations of individuals with reduced bias across demographic groups.
Topics
Interview‑Grounded Agents Achieve Near‑Human Consistency
Agents built solely from two‑hour semi‑structured interview transcripts recover ~82‑86% of participants’ test‑retest consistency on GSS items and outperform demographic baselines.
Robustness Analysis Shows Resilience to Partial Interview Loss
Systematic removal of 20‑80% of interview utterances leads to gradual declines in normalized accuracy, confirming that even truncated interviews retain valuable signal.
Comprehensive Interview Script as Rich Narrative Data Source
A two‑hour semi‑structured interview covering life story, health, finances, politics, values, and social experiences provides the primary grounding material for agents.
Stratified Sample Design and Quota Targets
The study employs a stratified U.S. sample with quotas on age, gender, region, race/ethnicity, education, and partisanship to ensure demographic representativeness.