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

Hanzhong Zhang; Muhua Huang; Jindong Wang

cs.AIAug 24, 2025

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

Authors: Hanzhong Zhang, Muhua Huang, Jindong Wang

Organizations: William & Mary, Williamsburg, VA, USA · Stanford University, Stanford, CA, USA

Abstract

Computational social science has developed descriptive, interpretative, and predictive modeling, but comprehensive modeling combining in-world human intervention with outcome forecasting remains underexplored. Generative agent-based platforms scale populations or reproduce complex scenarios, but researchers typically act through setup, external controls, or post-hoc analysis rather than in-world roles. We introduce Computational Multi-Agent Society Experiments (CMASE), a generative agent-based framework for in-world researcher participation and traceable interventions at runtime. Treating simulation as a simulated ethnographic field, CMASE supports (1) real-time human-computer interaction to study intervention processes; (2) controlled experiments with ethnography-inspired observation, interviewing, and interpretation; and (3) forward exploration of intervention trajectories and downstream responses. CMASE combines LLM-driven perception, memory, and emotion; a composable environment built from texture and area maps; and researcher roles spanning observation, interviewing, participation, and intervention, so that interventions enter as social events rather than exogenous parameters. We replicate two published field studies from different disciplines: urban greenery and social fragmentation, and cooperation in a public goods game, using two substantially different language models. CMASE reproduces reported statistical patterns and provides process-level records for mechanism-oriented interpretation. A controlled comparison shows that embedded social interventions leave observable, auditable processes that parametric manipulation does not. These findings highlight CMASE's methodological value for studying interventions in small- and medium-scale social settings and its interdisciplinary potential. The official code is available at: https://github.com/armihia/CMASE.

Figures & tables

Framework	Runtime human embedding	Intervention as social variable	Process tracing of interventions	Interpretive output
AgentSociety ( Piao et al., 2025 )	(✓) (external controls)	✗	(✓) (logged outcomes)	(✓) (interviews, surveys)
OASIS ( Yang et al., 2024 )	✗	✗	(✓)	✗
GenSim ( Tang et al., 2024 )	(✓) (global intervention)	✗	(✓)	(✓) (interviews)
Socioverse ( Zhang et al., 2025 )	(✓) (interview module)	✗	(✓)	(✓) (interviews)
Collaborative Gym ( Shao et al., 2024 )	✓ (task-level)	✗	✗	✗
HABITAT 3.0 ( Puig et al., 2023 )	✓ (embodied task)	✗	✗	✗

Table 1. Comparison of CMASE with representative GABM platforms and human–AI collaboration frameworks along the three capabilities required by comprehensive modeling and the availability of interpretive (ethnographic) output. ✓ = supported; (✓) = partially supported; ✗ = not supported.

Sec.	Evaluation	CMASE capability	Comparison or RQ	Principal measures and key outcomes
Study 1: Green Space and Social Fragmentation
4.3	Quantitative reconstruction	Pattern reproduction	Does CMASE reproduce the reported vegetation effect?	Indicator Low High Distrust ∗ 4.60 3.60 Exploitation ∗ 4.30 3.40 Indifference 3.70 2.70
4.4	Embedded inquiry	In-world inquiry	Can embedded inquiry reveal process-level patterns behind the aggregate result?	Indicator Observed pattern Indoor-agent coping Occupation-dependent coping Green-space use Lakeside clustering
4.5	Agent and record evaluation	Fidelity and interpretability	Are assigned personas preserved and do the records support interpretation?	Indicator Mean score Persona fidelity 4.23 Interpretive affordances 4.25
4.6	Runtime usability	Real-time participation	How does runtime scale, and what is the real-time participation bound?	Indicator Outcome Step time (1–10 agents) 23 s Real-time participation bound $\leq 100$ agents
Study 2: Public-Goods Cooperation and Intervention

Table 2. Evaluation roadmap of CMASE. Each evaluation is mapped to the CMASE capability examined, the comparison or research question, and its principal outcome measures and key findings.

	CMASE		Human NDVI		Human EVI
Dimension	Low Veg	High Veg	Low Veg	High Veg	Low Veg	High Veg
Distrust	$4.60\pm 0.97$	$3.60\pm 1.17$	4.0	1.6	4.5	1.1
Exploitation	$4.30\pm 1.34$	$3.40\pm 1.43$	5.0	1.4	4.5	1.6
Indifference	$3.70\pm 1.16$	$2.70\pm 1.49$	5.1	1.3	5.5	0.9

Table 3. Comparison of CMASE simulation results with human baseline.

Figure 2. Comparison of CMASE simulation with the original study results. (a) Simulation results generated by the CMASE framework. (b) Predictions using the EVI index from the original study. (c) Predictions using the NDVI index from the original study. (d) Movement trajectories of agents during CMASE simulation. (e) Regional variation in agents’ emotional states.

Persona Cluster	Consistency	Responsiveness	Emotional Alignment	Overall
Creative and Design	4.00	3.33	4.67	4.00
Analytical and Academic	4.33	3.33	4.33	4.00
Technical and Strategic	4.00	4.00	5.00	4.33
Public and Healthcare	5.00	4.50	5.00	4.83
Average	4.30	3.70	4.70	4.23

Table 4. LLM-based automated evaluation of Agent Persona Fidelity

Persona Cluster	Thick Description	Traceability	Interaction Depth	Mechanistic Insight	Overall
Creative and Design	4.33	4.00	3.33	3.67	3.83
Analytical and Academic	5.00	4.00	4.33	4.00	4.33
Technical and Strategic	4.00	4.00	4.50	4.50	4.25
Public and Healthcare	5.00	5.00	4.50	4.50	4.75
Average	4.60	4.20	4.10	4.10	4.25

Table 5. LLM-based automated evaluation of Ethnographic Affordances

Figure 3. CMASE performance under different agent counts. (a) Duration per time step. (b) Human evaluator responsiveness ratings.

Figure 4. Sealed condition. (a) Group trajectories coloured by composition, showing two persistent bands that rarely cross. (b) Distribution of endpoint cooperation across groups, coloured by composition, with defection-prone and balanced compositions occupying separate regions. (c) First-period contribution count against endpoint cooperation, with a Spearman correlation of 0.90.

Figure 5. Open condition. (a) Endpoint cooperation in the sealed and open conditions, paired by composition, with the defection-prone group falling to zero. (b) Mean trajectories for both conditions, with the open condition starting lower and falling faster. (c) Share of the per-period action budget by category, showing the move toward walking home and away from going to the box.

Figure 6. Delivery channels compared. (a) Mean group A trajectories for the baseline (O), the demonstration-only control (D), parametric injection (P), and embedded participation (E), with the intervention window and the point of withdrawal marked. (b) Mean cooperation in the intervention and released windows. The visible example contributes a gain of 1.93, while neither message channel adds anything beyond it, and the two channels yield the same mean intervention-window cooperation in these runs.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter	Value
Number of agents	10
Simulation length	50 time steps
Base model	GPT-4o (2024-12-01-preview)
Vegetation conditions	high / low (texture and perceptual layers)
Isolation mechanism	standing wall barriers with boundary guard
Gender distribution	50% male, 50% female

Appendix

Table 6. Parameter settings for the green-space community study.

Name	Occupation	Age	Long-term goal (abbreviated)
Joon-ho Kim	Software Engineer	31	Face nighttime panic alone and curb overreliance on community surveillance tools.
Min-jae Lee	Cardiologist	42	Accept his insomnia and helplessness, ease resentment toward the medical system.
Hyun-woo Park	University Professor	50	Stop letting formulas conceal inner emptiness; come to terms with social division.
Seok-jin Choi	Film Director	38	Put down the camera and allow himself to rest, not rely on the lens for connection.
Dong-hyun Jung	Entrepreneur	29	Let go of the startup-zealot persona during isolation.
Ji-eun Kim	Fashion Designer	33	Rediscover the purpose of her art beyond designing fashionable masks.

Appendix

Table 7. Agent personas and long-term goals for the green-space study. Goals are abbreviated; the full text is in the released configuration.

Computational Social Science

Computational social science uses data and computational models to study social behavior, institutions, communication, and collective outcomes.

See all 165 papers

Jun 11, 2026

AgentSociety 2: An Integrated Research Environment for Executable Social Science

AI scientist systems are beginning to automate parts of scientific research, but social science poses a distinct challenge: its objects of inquiry are not merely datasets or laboratory protocols, but integrated social processes involving situated participants, interaction contexts, interventions, and outcomes. Yet a critical link is missing: existing systems either assist isolated research tasks or simulate agents as experimental subjects, leaving the research workflow and simulated society decoupled. Here we introduce AgentSociety 2, an Integrated Research Environment for executable social science. It couples two roles of LLM agents in the same runtime: AI social scientists that coordinate literature grounding, hypothesis generation, experiment design, simulation execution, result interpretation, and manuscript drafting; and silicon participants that generate behavioral responses within configurable social environments. This dual-role design turns hypotheses into auditable agent behaviors, environment rules, interventions, and measurements, thereby supporting an end-to-end workflow. Across seven illustrative studies spanning micro-level social-science laboratory experiments, meso-level dynamics in social media, and macro-level urban scenarios, we demonstrate its capacity to support diverse disciplinary questions, reproduce major qualitative patterns from prior studies, identify informative deviations, and enable large-scale simulations through optimized agent-environment interactions. By preserving human researchers' high-level agency while delegating procedural orchestration to agentic systems, it provides a human-in-the-loop and controllable infrastructure for next-generation computational social science, with broader applications in scalable computational social experimentation and AI-enabled social governance platforms.

Jinghua Piao, Jun Zhang, Haoyu Huang +16

Department of Electronic Engineering, Tsinghua University · School of Environment, Tsinghua University · School of Public Policy and Management, Tsinghua University +2

Multi-Agent LLM Systems AI Agents for Scientific Discovery Human Behavior Simulation Computational Social Science Large Language Model-Based Social Simulation

Ask PDF

Jul 7, 2026

AgoraSim: A Hybrid Agent-Based Modeling Framework

LLM-agent simulations make natural-language social scenarios easy to instantiate, but their outputs can be overread as predictions and are often difficult to compare with explicit social dynamics. We present AgoraSim, a hybrid agent-based modeling framework for scenario-oriented social reaction analysis. AgoraSim resolves textual or multimodal artifacts into editable ABM configurations, runs ratio-controlled populations that mix LLM, vision-language, custom-endpoint, random, and classical agents, and compares the same scenario against matched classical reference dynamics. All agents emit a shared structured decision object, enabling common action spaces, interaction protocols, metrics, and audit records. Exposed through a local UI, Python SDK/CLI, and REST API, AgoraSim helps users inspect scenario trajectories, compare modeling assumptions, and identify cases that warrant empirical validation.

Chung-Chi Chen

Human-Agent Ally Lab (HAA Lab) National Institute of Informatics, Japan

Human Behavior Simulation LLM Agent Orchestration Agent-Based Modeling LLM Agents Large Language Model-Based Social Simulation

Ask PDF

Sep 21, 2026

SocioVerse2: A Longitudinal Dynamic Social Simulation Framework under a Human-AI Co-evolutionary Paradigm

Social simulation offers the social sciences an experimental instrument that the real world cannot supply, and generative agents have transformed it by acting as silicon samples that unite agent-based modeling with real behavioral data. Existing platforms verify collective behavior, align simulated populations with real societies in cross-sections, and employ autonomous agents for the research process. However, two social science requirements remain without systematic support: intervention in the content of a simulation and the researcher's control over the process that produces it. We present SocioVerse2, which extends SocioVerse 1.0 into a human-AI co-evolutionary paradigm built from two loops and one infrastructure. The longitudinal simulation loop simulates the target population with evolving environments and forks counterfactual branches via interventions. The controllable research loop takes the study itself as an editable state and updates state versions via controllable editing. The social science agentic infrastructure carries both loops through composable skills with researcher checkpoints, a population service over five persona pools, and an environment service over 21 real-world signal sources with point-in-time guarantees. We validate SocioVerse2 across three case families and seven case studies, from reproducing canonical agent-based models to modeling policy processes on real records and nowcasting macro-economic indices beyond the response model's knowledge cutoff. With the human-AI co-evolutionary paradigm, these cases go beyond system demonstrations to become substantive studies that investigate frontier questions in their respective disciplines. Code, data services, and a workbench are released as open-source resources.

Xinnong Zhang, Jiayu Lin, Jia Wang +19

Shanghai Innovation Institute · Fudan University · Tongji University +7

Computational Social Science Agent-Based Modeling Large Language Model-Based Social Simulation Human-AI Co-Creation

Ask PDF

Abstract

Computational Social Science

Explore similar work

AgentSociety 2: An Integrated Research Environment for Executable Social Science

AgoraSim: A Hybrid Agent-Based Modeling Framework

SocioVerse2: A Longitudinal Dynamic Social Simulation Framework under a Human-AI Co-evolutionary Paradigm