Can Low-Parameter LLMs Provide Trustworthy Simulations of Human BDI-II Responses? A Comparison of Biodemographic and Instagram-Based Context Profiles

This study evaluates whether publicly available low-parameter large language models can faithfully simulate human responses to the Beck Depression Inventory-II (BDI-II). It compares prompting strategies using personas constructed from biodemographic information or Instagram posts, with real participants' questionnaire responses serving as ground truth. The analysis examines both response accuracy and the preservation of individual differences. Results vary across models and experimental settings: some produce overly uniform representations, while others, particularly with Instagram context, capture more nuanced differences. The findings identify limitations that must be addressed before these models can reliably simulate human participants in psychological research.