Abstract
Large language models may support virtual patients. Most systems use manually prepared vignettes. We test a bottom up method that turns anonymized social media histories into virtual personas. GPT-4o and DeepSeek V4 Pro completed depression and anxiety scales across repeated generations. Depression scores had high single generation stability across conditions, with intraclass correlation coefficients from 0.84 to 0.88. Anxiety stability varied by model. DeepSeek V4 Pro had the highest alignment with a psychologist for depression under the Base prompt (r = .98). DeepSeek V4 Pro also had better item level separation. The results support bottom up persona construction. The results also indicate that evaluation must account for the model, prompt, and clinical scale.
Citation
Francesco Quilghini, Federico Torrielli, Amon Rapp, Luigi Di Caro, Michele Settanni, and Davide Marengo, “Developing Virtual Personas from User Level Social Media Data,” Preprints.org, 2026. DOI: 10.20944/preprints202607.1828.v1
@article{quilghini2026virtualpersonas,
title = {Developing Virtual Personas from User Level Social Media Data},
author = {Francesco Quilghini and Federico Torrielli and Amon Rapp and Luigi Di Caro and Michele Settanni and Davide Marengo},
year = 2026,
journal = {Preprints.org},
doi = {10.20944/preprints202607.1828.v1},
url = {https://doi.org/10.20944/preprints202607.1828.v1}
}