Computational Multi-Agent Society Experiments: Social Modeling Framework Based on Generative Agents

Computational social science has extensively developed descriptive, interpretative, and predictive modeling, but comprehensive modeling, which both admits human intervention into a simulated society and aims to forecast outcomes, remains underexplored. Existing generative agent-based platforms scale up populations or reproduce complex social scenarios, but researcher involvement is generally implemented through initial configuration, external runtime controls, or post-hoc analysis rather than as an explicit in-world role within the ongoing simulation. This paper introduces Computational Multi-Agent Society Experiments (CMASE), a generative agent-based framework that explicitly integrates the researcher into simulation runtime through role-based, in-world participation and traceable intervention channels. By treating the simulation as a simulated ethnographic field, CMASE shifts the researcher from an external operator to an embedded participant, and is designed to achieve three core capabilities: (1) enabling real-time human-computer interaction that allows researchers to dynamically embed themselves into the system to characterize complex social intervention processes; (2) combining controlled computational experimentation with ethnography-inspired observation, interviewing, and interpretation within the simulated field; and (3) supporting forward exploration of intervention trajectories and their downstream social responses within the simulation. Methodologically, CMASE realizes these capabilities through agents whose perception, memory, and emotion are driven by a large language model, a composable environment built from a texture and area map, and a typology of researcher roles spanning observation, interviewing, embedded participation, and intervention, so that an intervention enters the society as a social event rather than an exogenous parameter. We evaluate the framework through computational replications of two published field studies from different disciplines, one on urban greenery and social fragmentation and one on cooperation dynamics in a public goods game, using two substantially different language models. CMASE reproduces reported statistical patterns and yields process-level records that support mechanism-oriented interpretation within the simulated settings, while a controlled comparison shows that an intervention delivered as an embedded social event leaves an observable, auditable social process that a parametric manipulation does not. These findings demonstrate CMASE’s methodological value for studying intervention processes in small- and medium-scale social settings, highlighting its potential to advance interdisciplinary integration in the social sciences. The official code is available at: https://github.com/armihia/CMASE.