Title: Adaptive Querying with AI Persona Priors

Abstract:

This paper investigates adaptive querying strategies aimed at estimating user-specific metrics, such as responses to unseen items and psychometric traits, under strict query constraints. Traditional approaches, including classical Bayesian experimental design and computerized adaptive testing, are often hindered by rigid parametric assumptions or the computational cost of posterior approximations. These limitations restrict their applicability in environments characterized by high dimensionality, heterogeneity, and cold-start scenarios. To address these challenges, we propose a latent variable model driven by AI personas, which characterizes a user’s state via their affiliation with a finite set of persona types. Each persona is associated with response distributions generated by a large language model, resulting in expressive priors that support closed-form posterior updates and efficient finite-mixture predictions. This framework facilitates scalable Bayesian design for sequential item selection. Our evaluation on both synthetic datasets and WorldValuesBench reveals that persona-based posteriors yield precise probabilistic forecasts and support an interpretable adaptive elicitation workflow.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC