An interviewer–evaluator–judger LLM framework for text-based personality inference

This study presents an Interviewer–Evaluator–Judger framework that leverages large language models (LLMs) to assess individuals’ Big Five personality traits from written text, aiming to bridge computational modeling and human-centered personality understanding. Unlike prior approaches that rely on static, predefined personas, our method introduces an interviewer phase that dynamically constructs interview-based personas grounded in a user’s linguistic expressions. By retrieving texts similar to the input and prompting the LLM to elaborate and self-describe from a second-person perspective, the system builds a rich representation of personality cues that guide subsequent evaluation and judgment. We validate the method on three benchmark datasets—Essays, myPersonality, and Big-Five Backstage—and show that our framework generally improves performance over baseline and persona-based models, although the magnitude and consistency of the gains vary across datasets and LLMs. Results further reveal that model effectiveness depends on text origin, length, and contextual richness, suggesting that personality inference in human–AI interaction benefits from detailed, context-grounded persona representations. The findings provide insights into how LLMs can interpret human individuality through language and contribute to the development of more psychologically aware, trustworthy, and adaptive AI systems.