From Five to Six: Testing Cross-Inventory Generalization of LLM Personas Across Big Five and HEXACO Frameworks

Recent research indicates that Large Language Models (LLMs) can simulate human survey responses with moderate accuracy. It remains unclear whether these models can generalize personality profiles across theoretically distinct frameworks like transferring the Five-Factor Model (NEO-FFI) into the HEXACO-60 structure. This translation poses the psychometric challenge of recovering the Honesty-Humility factor, which is structurally absent in the NEO-FFI but latently distributed across the inventory. This study utilizes a within-subjects design with a human sample of N=1261 who participated in both the NEO-FFI and HEXACO 60 inventories. I condition an LLM on the NEO-FFI results of each participant to generate synthetic HEXACO responses under three experimental conditions of increasing granularity: - Domain-Prompt: Inputting the 5 NEO Domain Means. - Facet-Prompt: Inputting the 5 Domains + 13 Facets. - Item-Prompt: Inputting the full 60-item Likert response vector (methodology adapted from Huang et al., 2024) The focus is to investigate the proposition that increasing prompt granularity (from domains to items) significantly improves the fidelity of cross-inventory generalization. Expected outcomes are: - Outcome 1 (Accuracy): I expect the Item-Prompt condition to yield significantly higher convergent validity correlations with human HEXACO scores compared to the Domain-Prompt condition. - Outcome 2 (Structure): I expect high-resolution prompts to better preserve the internal factor structure (assessed via Tucker’s Congruence).