人工知能学会論文誌 人工知能学会設立40周年記念論文に当研究室の論文が採録

■書誌情報
森 隆太郎, 原田 憲旺, 水野 晋之介, 大西 直, 堀口 維里優, 坂本 航太郎, 堀部 和也, 小島 武, 松尾 豊, 岩澤 有祐 (*Equal Contribution): 大規模言語モデル時代の社会シミュレーション「再現」と「理解」の目的別に見る研究動向, 人工知能学会論文誌 (JSAI 2026), 人工知能学会設立40周年記念論文, 2026
■概要
This paper surveys social simulation research using Large Language Models (LLMs), arguing that their contributions should be evaluated through two distinct objectives: (1) building a simulation as a faithful replication of a specific real-world system, and (2) building a simulation as a theoretical model of mechanisms that recur across multiple systems. Social simulation has historically pursued these aims separately: constructing detailed models of particular systems (e.g., organizations, cities) to test interventions, and developing abstract toy models to identify general principles (e.g., polarization, segregation). The “human-like” capabilities of LLM-based agents hold promise for both directions, yet their value and challenges differ fundamentally. For the purpose of faithfully replicating specific systems, LLMs enable agents to embody complex individual attributes and engage in natural language-based interactions difficult to model traditionally. The central challenge is validation and calibration against real-world data to yield trustworthy insights for decision-makers. For the purpose of working as theoretical models, LLMs expand the scope of theoretical modeling by internalizing linguistic interactions (e.g., negotiation, persuasion, rule modification) within agents, making phenomena such as opinion propagation, polarization, and norm formation amenable to formal analysis. A longer-term direction within this objective may be “high-dimensional science,” where detailed multi-system simulations combined with large-scale predictive models and interpretability techniques may extract general structures in a data-driven manner. Key challenges on the generality path include clarifying the theoretical status of LLM agents and developing methods to extract interpretable principles from complex simulations. Overall, this survey argues that explicitly separating these two classical objectives enables the field to define appropriate evaluation criteria and research agendas for each.