原著論文
大規模言語モデルを用いた定量的ペルソナの生成・運用フレームワーク
キーワード:
Quantitative Persona,
Social Simulation,
Large Language Model,
Retrieval-Augmented Generation,
Reasoning,
Acting
2026 年 22 巻 1 号 p. 1-7
詳細
2026 年 22 巻 1 号 p. 1-7
This paper describes a constructive architecture for using large language model (LLM) in generating virtual personas with respect to the authors' proposed simulation-based quantitative persona creation method. This architecture consists of a layered structure for persona generation and the concept of operational evolution. The layered structure consists of a Fundamental Layer consisting of typified facts extracted from simulations, a Narrative Layer generated from facts typified by LLM, and a Presentation Layer that uses LLM to enhance the realism of the personas as personas. In the operational phase, this layered structure allows for easy modification and reflection based on feedback from the field and changes in the environment. The main findings of this study are as follow: our architecture could 1) achieve logical consistency with simulation and survey results, 2) generate stories with a certain paragraph structure, 3) inject external information and knowledge, and 4) derive the technical process naturally to generate the personas. In particular, we conduct specific demonstrations about 3) above mentioned; we found that it is possible to revise persona stories in a realistic manner by adding external information to them through prompt engineering methods like Reasoning and Action; ReAct.
