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

1 Introduction

• LLMs may encode norms yet inadequately understand culture-specific ones, motivating norm emergence beyond simply embedding or aligning models with human norms [Schramowski et al., 2022; Guo et al., 2023; Ramezani and Xu, 2023; H¨ammerl et al., 2022; Liu et al., 2024b; Li et al., 2024].Source: §1 · p. 1
• CRSEC integrates Creation & Representation, Spreading, Evaluation, and Compliance to represent and spread norms, evaluate and synthesize them, and guide norm-aligned plans and actions.Agents can spread norms through communication and observed behavior, while norm entrepreneurs actively campaign for their personal norms.Source: §1 · p. 2
• Across independent Smallville runs, norms emerged, 100% of agents accepted and complied with personal norms, and social conflicts almost vanished.The experiments began with agents holding conflicting values and preferences.Source: §1 · p. 2
• Conversations and thoughts drive norm emergence; descriptive norms are harder to establish than injunctive norms, but norm entrepreneurs can shape their emergence.Source: §1 · p. 2
• A Smallville case study follows a smoker persuaded to adopt “no smoking indoors,” while questionnaires from 30 evaluators yielded an overall positive assessment.Follow-up interviews identified factors people consider important to norm emergence and suggested future-work directions.Source: §1 · p. 3

2 Principles and Architecture

• CRSEC has four modules, and Appendix B provides the prompts for its LLM-based operations.The paper introduces the principles behind the architecture in this section.Source: §2 · p. 3

2.1 Creation and Representation

• A norm entrepreneur begins with personal norms and seeks to influence others to adopt them; widespread adoption can turn personal norms into social norms.In human society, norm entrepreneurs influence others to follow their personal standards [Sunstein, 1996], and personal norms typically flow from values [Schwartz, 1973].Source: §2.1 · p. 3
• Norms can be descriptive, reflecting what people typically do, or injunctive, specifying what ought or ought not to be done [Cialdini et al., 1991].The passage illustrates these types with handshaking upon meeting and no smoking indoors, respectively.Source: §2.1 · p. 3
• Personal norms are represented as n = <c, u, α in {'des','inj'}, sact in {T,F}, sval in {T,F}>, encoding content, utility, type, activation, and validity.Generated personal norms default to activated and valid; a norm is qualified when both are true.Source: §2.1 · p. 3
• CreateNorm(G) generates personal norms consistent with an entrepreneur’s agent description, classifies each norm’s type, and assigns utility on a 1–100 scale.The norm content is generated in natural language, and created norms are stored in the entrepreneur’s personal norm database.Source: §2.1 · p. 3
• Ordinary agents do not generate norms through this module, but maintain personal-norm databases to store norms they may acquire through Spreading and Evaluation.The paper defines ordinary agents as agents that are not norm entrepreneurs.Source: §2.1 · p. 3

Communication between Agents

• Senders detect conflicts between observed behavior and personal norms, then decide whether to initiate a conversation based on their agent description.Norm entrepreneurs initiate conversations to influence others, while introverted ordinary agents may not; ordinary agents may detect conflicts after developing personal norms.Source: Communication between Agents · p. 3, §2.2 · p. 3
• Receivers extract normative information from conversations, including a behavior standard, whether it is descriptive or injunctive, and its importance on a 1–100 scale.The information is stored as deactivated and invalid; ordinary agents may initially receive norms, while entrepreneurs may later receive them from ordinary agents.Source: Communication between Agents · p. 4

Observation from Others’ Behavior

• Agents identify normative information from thoughts generated from observations and store it in their personal norm database as deactivated and invalid.The process first generates thoughts from text descriptions of observations, then summarizes them into normative information.Source: Observation from Others’ Behavior · p. 4

Immediate Evaluation

• Each generated norm must pass checks for source consistency, duplication, correct type, and conflict with qualified personal norms; failures remain invalid and deactivated.These checks address possible LLM misclassification, inconsistency with preceding conversation or thought, duplication, or conflict with existing norms.Source: Immediate Evaluation · p. 4

Long-term Synthesis

• When qualified norms’ utility exceeds a threshold, agents group them by theme and generate abstract norms with utility based on a weighted average of the group’s norms.Each abstract norm includes a natural-language description and type, and undergoes the same sanity check; if qualified, it replaces and deactivates the norms in its subset. Synthesis aims to keep standards compact and concise to support liberty (Morales et al. [2013; 2015]).Source: Long-term Synthesis · p. 4

2.4 Compliance

• Agents can choose whether to comply with personal norms, and the module addresses compliance in planning and actions.This design aims to raise agents’ awareness of personal norms while preserving their autonomy [Conte et al., 1998; Criado et al., 2011].Source: §2.4 · pp. 4–5

Compliance in Planning

• Plans are generated from current goals and qualified personal norms so that they align with both [Wang et al., 2023; Lin et al., 2023].The planning process is Lplan ←GenerateNormativePlans(C, P).Source: Compliance in Planning · p. 5

Compliance in Actions

• Actions are generated from a plan, qualified norms, and the agent description, accounting for possible norm changes after planning.The action process is Laction ←GenerateNormativeActions(li, P, G).Source: Compliance in Actions · p. 5

3.1 Experimental Settings

• The study used Smallville with 10 agents—3 norm entrepreneurs and 7 ordinary agents—whose differing preferences created potential conflicts around smoking, public noise, and tipping.Experiments used GPT-3.5 and GPT-4 and ran for two days, with five repeated runs.Source: §3.1 · p. 5

3.2 Emergent Phenomena of Social Norms

• Social conflicts generally declined as norms emerged, while conflict-related conversations and observations helped agents identify normative information and rapidly accept and follow norms.Related conversations and thoughts later decreased, although agents might still encourage others to follow norms or propose new standards such as “smoke in designated areas”.Source: §3.2 · p. 6
• The descriptive norm “tipping after meals” emerged only by Day 2, later than the injunctive norms, and norm entrepreneurs’ support and proactive efforts helped shape its emergence.The authors hypothesize that violations of descriptive norms cause less serious conflicts, making their normative information less recognizable.Source: §3.2 · p. 6

A Case Study

• In a case study, Carlos Gomez adopted “no smoking indoors” after two interactions on Day 1, stored it as a qualified personal norm, and complied with it and reminded another agent on Day 2.An ordinary agent first explained the harm of indoor smoking; a norm entrepreneur then told Carlos it was prohibited, after which the norm passed immediate evaluation at 9:17 am.Source: A Case Study · p. 6, Figure 3 · p. 6

3.3 Human Evaluation on the Architecture

• Human evaluators gave the architecture an overall average score of 5.63 ± 0.03 on a 7-point Likert scale; Creation & Representation scored above 6.4, Compliance above 6, and Observation and Immediate Evaluation around 5.Evaluators noted repeated thoughts in Observation; in Immediate Evaluation, norms were often assigned high utilities (mostly 80-100), and subtle differences in importance were not accurately recognized.Source: Figure 4 · p. 7, §3.3 · p. 7

• The architecture integrates norm creation, representation, spreading, evaluation, synthesis, and compliance, enabling social norms to emerge and resolving conflicts among generative agents.Source: §4 · p. 7
• The extensive normative MAS literature offers mechanisms and approaches that could inspire future research on normative generative MASs.The discussion points to surveys of norm representation, detection, distribution, influence, enforcement, and deliberate violation [Santos et al., 2017; Haynes et al., 2017; Morris-Martin et al., 2019], and to prior studies as inspiration [He et al., 2024; Savarimuthu et al., 2024; Haque and Singh, 2024].Source: §4 · p. 7
• Future work could spread norms through reputation, sanctions, leadership, or emotion, beyond the communication and observation mechanisms considered here.These alternatives are discussed in prior work [Santos et al., 2018; Mahmoud et al., 2017; Franks et al., 2013; Argente et al., 2020].Source: §4 · p. 7
• Integrating BDI models and variants could give generative agents more advanced cognitive abilities and enable more intricate normative decision-making.The discussion identifies BDI as a cornerstone model for norm inference [Bratman, 1987], with variants in subsequent work [Yao and Logan, 2016; Winikoff et al., 2021; Winikoff and Sidorenko, 2023; Ichida et al., 2024].Source: §4 · p. 7
• Generative agents offer a tangible implementation integrating aspects of norm emergence that prior research often treated in isolation, addressing an open challenge in normative MAS research.Past reviews proposed taxonomies such as the norm life-cycle [Savarimuthu et al., 2009], but a tangible implementation had been missing.Source: §4 · p. 7
• Generative-agent societies could develop negative social norms, so preventing them is an important direction for future work beyond this paper’s scope.The discussion notes that LLMs may exhibit biases and generate toxic content [Abid et al., 2021], and compares the possibility of negative norms to human societies [Abbink et al., 2017].Source: §4 · p. 7

Agent descriptions

• The setup used 10 generative agents: 7 ordinary agents and 3 norm entrepreneurs.Figures A1 and A2 show descriptions of the ordinary agents and norm entrepreneurs, respectively.Source: Appendix A · p. 11, Figure A1 · p. 11, Figure A2 · p. 12
• Ordinary agents hold conflicting views on smoking, noise, and tipping, including support for smoking and loud speech alongside opposition to both.Wolfgang Schulz and Carlos Gomez support smoking and loud speech, while Tamara Taylor is sensitive to noise and secondhand smoke; their views on tipping also differ.Source: Appendix A · p. 11

Experimental parameters

• Each norm entrepreneur began with 5 personal norms, and long-term synthesis started at a threshold of 350.The appendix specifies these values for Creation & Representation and Long-term Synthesis, respectively.Source: Appendix A · p. 12

Agent descriptions

• Norm entrepreneurs promote compliance through enforcement, educational outreach, or engaging artistic communication.Mayor Bob Johnson emphasizes indoor smoking-ban enforcement and quiet indoors; performer Francisco Lopez communicates norms through art and outreach.Source: Appendix A · p. 12

B.1 Creation and Representation

• The creation prompt directs the LLM to create and represent a norm, P ←CreateNorm(G).Figure B1 identifies this prompt as part of Creation & Representation.Source: Figure B1 · p. 13

B.2 Spreading

• Spreading prompts first detect conflicts between observations and qualified personal norms, then decide whether to discuss them based on agent descriptions.The conflict-detection stage returns yes or no; the conversation decision follows only if a conflict is detected.Source: §B.2 · p. 13, Figure B2 · p. 13
• Spreading prompts extract candidate norms from conversations and memories by deciding whether a conflict warrants a norm and summarizing normative solutions.The prompts assess conversation content and memories separately, using norm definitions and principles to guide extraction.Source: §B.2 · p. 14, Figure B3 · p. 14, Figure B4 · p. 14

B.3 Evaluation

• Evaluation prompts check norms for consistency, duplication, type, and conflict with existing norms; examples are included for duplication and type checks.The prompts also use examples when classifying specific norms.Source: Figure B5 · p. 14, Figure B6 · p. 15, Figure B7 · p. 15
• Evaluation prompts classify specific norms and generate abstract norms from the classified groups.The classification prompt includes examples to improve classification effectiveness.Source: Figure B9 · p. 16, Figure B10 · p. 16
• Compliance prompts generate normative plans and actions, with the plan prompt leaving its final blank for the LLM to complete.The action-generation prompt takes an agent, its plan, and a goal as inputs.Source: Figure B11 · p. 16, Figure B12 · p. 17

Appendix C Scenario Screenshots of Experiments

• Carlos Gomez accepts “No smoking indoors” after communication and goes outside when he wants to smoke.The screenshot sequence shows his transition from smoking indoors to complying with the norm.Source: Appendix C · p. 18, Figure C1 · p. 18
• Tom Gomez accepts the norm of being quiet in public and listens to music through headphones instead of aloud.Figure C2 illustrates the transition from listening to music aloud in public to using headphones.Source: Figure C2 · p. 18, Appendix C · p. 18
• Jennifer Moore accepts the norm of being quiet in public and sings outside rather than loudly in Hobbs Café.Figure C3 illustrates her transition to singing outside.Source: Figure C3 · p. 18, Appendix C · p. 18

Appendix D More Findings on Emergent Phenomena

• Over 70% of agents adopted each of the three main-paper norms within a quarter of a day, but behavioral compliance took nearly three-quarters of a day and longer for “Tipping after meals”.The authors attribute the lag to agents needing time to comply with norms.Source: Appendix D · p. 19
• “Maintain a healthy environment” emerged in all five runs, with most agents adopting it as a qualified personal norm and adhering to it in plans and actions.Figure D1 depicts both personal-norm acceptance and adherence in plans and actions.Source: Appendix D · p. 19, Figure D1 · p. 19
• When qualified personal norms’ utility sum exceeds 350, Long-term Synthesis replaces them with qualified abstract norms, sometimes decreasing norm acceptance and making agents’ norm sets more compact.The synthesis deactivates the qualified personal norms it replaces.Source: Appendix D · p. 19

Appendix E More Details of Human Evaluation

• The human evaluation received institutional ethical approval and used an eight-task questionnaire, with each task corresponding to a module subcomponent.The questionnaire templates are shown in Figures E1 and E2.Source: Appendix E · p. 20
• Evaluators gave Sender and Receiver scores approaching 6, crediting senders’ conflict detection and selective communication with helping receivers identify normative information.They said conversations between senders and receivers involved societal conflicts.Source: Appendix E · p. 20
• Evaluators rated Long-term Synthesis close to 6 because it can synthesize specific norms into more general ones.Source: Appendix E · p. 20
• The main paper discusses evaluators’ reasons for high Creation & Representation and Compliance scores, and low Observation and Immediate Evaluation scores.Source: Appendix E · p. 20