From Persona Specifications to Realized Behaviour in Medical Dialogue Simulation

Large language model (LLM)-powered simulated patients are increasingly used to evaluate and train medical dialogue systems, yet their behavioral controllability remains poorly understood. In this study, we investigated how persona specifications influence simulated patient behavior by comparing two contrasting profile modes—Silent Regular and Anxious Information-Seeker—across nested persona bundles containing 0, 5, 9, or 13 attributes across matched clinical vignettes. Our analysis revealed two primary findings. First, broader persona bundles consistently increased behavioral differentiation relative to case-only baselines, but this effect was non-monotonic: differentiation peaked at nine attributes and yielded no further gains at thirteen. Second, individual attribute controllability was highly uneven; high anxiety consistently drove localized behavioral shifts, such as increased reassurance-seeking, whereas altering disclosure style resulted in marginal and inconsistent changes. Together, these findings demonstrate that while persona breadth is an effective tool for behavioral differentiation, increasing specification richness does not invariably produce richer behaviors, highlighting the nuanced dynamics of simulated patient controllability.