SoNoLiSi: Simulating the Social Norm Lifecycle with Generative Agents

Rasika Muralidharan, Haewoon Kwak, Jisun An

Keywords: LLM, multi-agent, social simulation

Abstract:

Social norms play a crucial role in guiding the behavior of individuals within groups, reducing conflict and enabling coordination. Recent work using large language model (LLM) agents in generative agent-based models has demonstrated that normative behavior can emerge in multi-agent systems (Xu et al., 2024; Ashery et al., 2024; Jin et al., 2024). However, most studies focus on reproducing normative outcomes rather than investigating the mechanisms required for norm emergence.
In this work, we propose a framework for studying social norm emergence in language-based artificial societies through mechanistic ablation experiments. Interdisciplinary theories of social norms suggest that norms depend on shared expectations about what individuals ought to do and what they believe others will do, emerge through social learning, and are maintained through reputation, social inclusion, and exclusion (Bicchieri, 2006; Young, 2015). We operationalize these theories computationally using generative agents and examine three mechanisms commonly identified in norm theory: social learning, selection pressure, and oughtness (beliefs about what behaviors others expect).
Under the social learning mechanism, agents interact through discussion, enabling the communication of expectations, while social selection allows for reputation-based selection for agents to adjust future interaction partners based on past behavior. We systematically ablate these mechanisms across several experimental conditions, including a baseline without discussion or selection, discussion without selection, selection without discussion, and a full model including both mechanisms. Using repeated fairness game interactions, we analyze how discussion and reputation-based partner selection influence behavioral convergence, norm recognition, stabilization, and selective exclusion.
We evaluate these dynamics using outcome variables including cooperation rates, convergence of injunctive expectations (“oughtness”), network exclusion patterns, and time to behavioral stabilization.
Preliminary results suggest that discussion plays an important role in norm recognition, while partner selection contributes to the maintenance and stabilization of norms even in the absence of explicit sanctioning. In contrast, the baseline condition exhibits polarization and fragmented behavioral clusters, whereas the full model produces greater convergence toward cooperative norms and reduced behavioral variance.
For robustness, we further extend the framework to explore norm emergence in Wikipedia policy discussions, grounding the simulation in empirical communication data. Additionally, we test the model across various parameters including: a) group size (capped at 15 due to Dunbar’s number and the use of reputation mechanism), b) percentage of violators in a group, c) model families—GPT, Llama, Mistral and Qwen, d) each experiment is run 4 to 10, depending on cost.
This work demonstrates how generative agent simulations can serve as computational testbeds for evaluating theoretical mechanisms of social norm emergence. By enabling controlled experiments on learning, expectations, and social selection, this framework moves beyond outcome-based demonstrations toward mechanistic explanations of how norms arise and stabilize in language-mediated artificial societies.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 134