Emergence of Social Norms in Generative Agent Societies: Principles and Architecture

TL;DR

This work introduces CRSEC, a normative architecture for generative multi-agent systems that enables the emergence of social norms through four modules: Creation & Representation, Spreading, Evaluation, and Compliance. Personal norms created by norm entrepreneurs are stored, propagated via communication and observation, evaluated for consistency and novelty, and synthesized into concise abstract norms when appropriate, with compliance guiding planning and action. Empirical validation in the Smallville sandbox demonstrates robust emergence of injunctive norms, substantial reduction in social conflicts, and positive human assessments, reinforcing the viability of normative generative MASs. The study highlights practical implications for trust, coordination, and human-agent interactions, while outlining future work on extensions such as reputation, sanctions, and emotion in norm dynamics.

Abstract

Social norms play a crucial role in guiding agents towards understanding and adhering to standards of behavior, thus reducing social conflicts within multi-agent systems (MASs). However, current LLM-based (or generative) MASs lack the capability to be normative. In this paper, we propose a novel architecture, named CRSEC, to empower the emergence of social norms within generative MASs. Our architecture consists of four modules: Creation & Representation, Spreading, Evaluation, and Compliance. This addresses several important aspects of the emergent processes all in one: (i) where social norms come from, (ii) how they are formally represented, (iii) how they spread through agents' communications and observations, (iv) how they are examined with a sanity check and synthesized in the long term, and (v) how they are incorporated into agents' planning and actions. Our experiments deployed in the Smallville sandbox game environment demonstrate the capability of our architecture to establish social norms and reduce social conflicts within generative MASs. The positive outcomes of our human evaluation, conducted with 30 evaluators, further affirm the effectiveness of our approach. Our project can be accessed via the following link: https://github.com/sxswz213/CRSEC.

Figures (24)

• Figure 1: CRSEC: our architecture for the emergence of social norms in generative agent societies. Initially, by the Creation & Representation module, norm entrepreneurs create their personal norms and store them into their databases. By the Spreading module, some agents proactively influence others to adopt their personal norms through initiating communication with others, while others can identify those norms from their chats and observations. The identified norms then undergo an immediate evaluation in the Evaluation module. The Compliance module enables agents to generate plans and actions, with the norms bearing in mind. The normative actions, in turn, can influence other agents' observations and thus reinforce the spreading of norms. In addition, from time to time, agents perform long-term synthesis to keep their personal norms compact and concise.
• Figure 2: The evolution of generative MASs. Panel (a) depicts the evolution of the number of social conflicts, thoughts and chats over time. Panel (b) illustrates the emergent process of social norms in terms of (i) the proportion of agents that have accepted a standard of behavior as their personal norms in their databases, and (ii) the proportion of agents that have adhered to a standard of behavior in their plans and actions.
• Figure 3: A case study illustrating how a seasoned smoker has gradually adopted "no smoking indoors" as his personal norm.
• Figure 4: Human evaluation results. The overall averaged score of our architecture is $5.63\pm0.03$. Note that we use 7-point Likert scale, ranging from strongly disagree (1), disagree (2), somewhat disagree (3), neutral (4), somewhat agree (5), agree (6), to strongly agree (7).
• Figure A1: Agent descriptions of seven ordinary agents.