From Individual to Society: A Survey on Social Simulation Driven by Large Language Model-based Agents

1 Introduction

• Traditional sociological research uses questionnaires and psychological experiments, which can provide authentic data but are expensive, difficult to scale, and involve ethical risks.These methods are used to test theoretical hypotheses, understand social phenomena, and predict collective outcomes.Source: §1 · p. 1
• LLM-driven agents can simulate corresponding individuals’ responses through role-playing, helping researchers understand, validate, and predict human reactions.The paper calls this response-simulation property algorithmic fidelity.Source: §1 · p. 1
• Multiple LLM-driven agents can collaborate on specialized tasks or generate collective patterns that replicate complex social dynamics.The passage gives software development, diagnosis, and judicial decision-making as collaborative tasks, and opinion dynamics and macroeconomics as social phenomena.Source: §1 · p. 1

Society Simulation

• Individual simulation models a specific person or demographic group, focusing on features such as personality without multi-agent interaction.The figure presents the three types from left to right as generally increasing in diversity and scale, while individual-modeling granularity becomes more refined from right to left.Source: Society Simulation · p. 2, Figure 1
• Scenario simulation organizes agents around specific goals or tasks, usually in small groups with specialized expertise.Examples include software development, question answering, and paper reviewing.Source: Society Simulation · p. 2
• Society simulation models complex and diverse behaviors in agent societies to explore real-world social dynamics.It can examine social science theories at small scope or populate virtual spaces and communities with large-scale social phenomena.Source: Society Simulation · p. 2
• The three types progress from individual modeling toward broader simulations, with individual simulation serving as a foundation for scenario and society simulation.Society simulation could theoretically include countless subscenarios, although current work focuses on specific scenarios.Source: Society Simulation · p. 2
• The survey addresses the gap in systematic coverage from individual to society by reviewing architectures, objectives or scenarios, and evaluation methods across the three simulation types.It also covers datasets and benchmarks and analyzes trends across the types.Source: Society Simulation · p. 2
• An individual-simulation agent typically combines profile, memory, planning, and action modules, built through prompting or training for objectives such as simulating characters or demographics.Evaluation may be static or interactive and observe different dimensions.Source: Figure 2 · p. 3

2.1 Large Language Model-based Agents

• LLM-based agents incorporate memory, planning, and tool use to support effective operation in complex and diverse environments.Memory stores and operates on historical information, planning decomposes complex tasks into subtasks, and tools provide access to external resources.Source: §2.1 · p. 3

2.2 Multi-agent Systems

• Multi-agent systems involve agents that may share goals or pursue self-interested goals, with distinct roles, skills, and tasks.Agents can be organized in layered or centralized structures and communicate through different methods, factors that influence interaction effectiveness and efficiency.Source: §2.2 · p. 3

3.1 Architecture

• Individual-simulation agents integrate individualized data through profile, memory, planning, and action components, with profiles built from manual sources or LLM generation and represented as descriptions or conversations.Manual sources include handcrafted profiles, online communities, and historical works; descriptions convey basic attributes, while dialogue can reflect deeper context.Source: §3 · p. 3, §3.1 · p. 3, §3.1.1 · p. 5

3.1.2 Memory

• Memory stores short-term interaction content and persistent long-term information, while writing, retrieval, and reflection update and use those records.Reflection can refine memories into more abstract concepts, although its applications remain limited.Source: §3.1.2 · p. 5

3.1.3 Planning

• Individual-level planning extends goal-directed action with empathetic planning about others’ behavior and emotions and subjective planning based on the agent’s own identity and internal state.Empathetic planning uses reasoning to adapt decisions to others’ situations; subjective planning can use inner monologues or reasoning to express the agent’s own beliefs.Source: §3.1.3 · pp. 5–6

3.1.4 Action

• Action connects LLMs to their environment across situations from simple dialogue to crafted games or sandboxes, sometimes adding script-like expressions or environmental detail.Open-domain settings allow freer responses, whereas closed-domain settings restrict the available actions.Source: §3.1.4 · p. 6
• Closed-domain simulations constrain the action space to predefined choices or tools, while open-domain simulations impose few restrictions and more closely resemble real-world conditions.Closed-domain tasks can also extend to practical tool use, including recommending, browsing, or compiling.Source: §3.1.4 · p. 6

3.2 Construction

• Construction integrates individual-related data through nonparametric prompting or parametric training, which includes pre-training, fine-tuning, and reinforcement learning.Prompt-based simulations can be limited by context windows and preset assistant templates; fine-tuning uses task-specific data, while reinforcement learning refines behavior in dynamic settings.Source: §3.2 · p. 6, §3.2.1 · p. 6, §3.2.2 · pp. 6–7

3.3 Simulation Objectives

• Simulation objectives cover demographics, such as groups sharing traits or identities, and distinct real or fictional characters; demographic studies include surveys and preference or bias evaluation.Character simulations draw on public figures and fictional roles, though virtual-character work faces dataset quality and reliability challenges.Source: §3.3 · p. 7, §3.3.1 · p. 7, §3.3.2 · p. 7

3.4 Evaluation

• Evaluation ranges from static dialogue assessment by human or LLM judgments and objective metrics to interactive assessments in designed environments with real-time responses and multiple stages.Objective measures include accuracy, F1, recall, perplexity, ROUGE-L, and BLEU; the survey notes that objective examination is restricted and may require task-specific tools.Source: §3.4 · p. 7, §3.4.1 · pp. 7–8, §3.4.2 · p. 8

4 Scenario Simulation

• Scenario simulations organize multiple agents to interact and collaborate on tasks, including software development, question answering, and paper reviewing.The motivating question is whether agents can cooperate like humans or surpass human performance in achieving collective intelligence.Source: §4 · p. 9
• A scenario simulation typically designs a multi-agent system by specifying its environment, agent roles, organizational structure, and communication protocols.These components are adjusted to the targeted scenario and task.Source: §4 · p. 9

4.1.1 Environment

• An environment establishes scenario context through task events and agent profiles, and may provide agents with external resources or tools.Tools range from programming languages and structured-data queries to calculators and APIs.Source: §4.1.1 · pp. 9, 11
• Environment states include observations and action-dependent feedback, while accumulated history can be integrated, refined, summarized, or handled through memory mechanisms.Feedback can include changes to agents’ cognitive states or strategies and results from external tools.Source: §4.1.1 · p. 9

4.1.2 Role

• Participants execute tasks as communicators who exchange information or workers who perform specialized operations.Examples include coders and testers, negotiators, and medical professionals.Source: §4.1.2 · p. 11
• Directors support task execution as planners, coordinators, and integrators who respectively formulate plans, manage collaboration, and consolidate or decide outcomes.Examples include project managers, critic or judge agents, and summarizers.Source: §4.1.2 · p. 11

4.1.3 Organization

• Organizations may be static, with predefined single- or multi-stage arrangements, or dynamic, allowing agents to be recruited, coordinated, or self-organized as tasks unfold.Dynamic structures can also span multiple stages with communication shaped by the current state.Source: §4.1.3 · pp. 11–12
• Organizational connections can be layered, centralized around a coordinating role, or decentralized for peer-to-peer interaction.These structures describe how agents connect, distinct from whether their organization remains stable or changes during a simulation.Source: §4.1.3 · pp. 11–12

4.1.4 Communication

• Agents communicate through flexible but potentially ambiguous natural language or structured formats such as code and JSON, and communication may be cooperative or competitive.Cooperative settings pursue shared outcomes, whereas competitive settings involve differing positions and individual objectives.Source: §4.1.4 · p. 12

4.2.1 Dialog-Driven Scenario

• Social-interaction scenarios study tasks such as persuasion and comforting, with work also evaluating social intelligence and learning social skills.Studies examine differences across models and extend evaluation to objective action-level measures and reasoning about others’ information [138, 140, 204–207].Source: §4.2.1 · p. 12
• Question-answering scenarios use collaboration, debate, strategic reasoning, and negotiation to improve task performance.Examples include multi-agent debate for commonsense reasoning, mathematical reasoning, and factual accuracy, as well as systems targeting negotiation and reasoning reliability [144] [29] [148] [149].Source: §4.2.1 · p. 12
• Games progress from reproducing communication games to studying social dynamics, adaptive strategies, opinion leadership, and ad hoc teamwork.Reinforcement-learning frameworks are reported to achieve near-human-level decision-making [151] [155] [159] [160] [156] [157] [208].Source: §4.2.1 · pp. 12–13

4.2.2 Task-Driven Scenario

• Task-driven scenarios assign agents functional personas for specialized tasks, with applications in science, software development, and other industries.Examples include medical reasoning, chemistry research workflows, collaborative software lifecycles, legal decision-making, economic markets, and education [18, 166–168, 209, 167, 165, 176–180, 186, 181, 20, 182, 183, 187, 184].Source: §4.2.2 · p. 13

Task Evaluation

• Task evaluation measures overall scenario-task performance with automatic metrics, LLM judgments, or human assessment.Automatic measures are efficient and scalable but may miss complex behaviors; LLMs and experts can provide more nuanced qualitative comparisons.Source: §4.3 · p. 13

Sub-Task Evaluation

• Subtask evaluation tracks component completion and its effect on overall performance, combining process metrics with LLM-based or human judgments.Metrics include transport rate, steps, success, replanning, efficiency, and software completeness, executability, and consistency.Source: §4.3 · p. 13

System Evaluation

• System evaluation assesses whole-system effectiveness and efficiency, while the society-simulation framework also evaluates individual and system performance.Measures include token use, task success, human-likeness, diagnostic accuracy, runtime, and resource use.Source: §4.3 · p. 14, Figure 4 · p. 14

5 Society Simulation

• Society simulations investigate emergent interaction outcomes and social dynamics, supported by four construction dimensions: composition, network, social influence, and outcomes.The survey frames society simulation as useful for theoretical validation and predicting social dynamics, and summarizes representative works in Table 3.Source: §5 · p. 14, §5.1 · p. 14, Table 3 · p. 15

5.1.1 Composition

• Composition choices range from synthesized or dataset-based populations to real-world distribution replication, with detailed agent profiles traded off against simulation scale.Some studies prioritize detailed modeling of influential outliers while simplifying the majority.Source: §5.1.1 · pp. 14, 16

5.1.2 Network

• Social networks encode relations that guide information and influence flow; online networks may combine real relationships with synthetic links as populations grow.Other approaches use dataset or synthesized-user relations, or connect similar users under a homophily assumption.Source: §5.1.2 · p. 16

5.1.3 Social Influence

• Social influence depends both on recipients’ traits and on senders’ status, while larger groups can exhibit a herd effect.Profiles, memory, cognitive bias, norms, reputation, and social impressions are among the modeled factors.Source: §5.1.3 · pp. 16–17

5.1.4 Outcomes

• Outcomes include aggregate statistics such as opinions or votes and emergent phenomena or norms such as echo chambers, cooperation, and regulatory effects.Some studies sum individual choices directly, while others model multiple interaction rounds before analyzing changes.Source: §5.1.4 · p. 17

5.2.1 General Economics

• Economic simulations examine resource allocation, competition, strategic interaction, and macroeconomic behavior under incentives, rules, or constraints.Classic games support controlled studies, while real-world scenarios can offer more application-oriented insight.Source: §5.2 · p. 17, §5.2.1 · p. 17

5.2.2 Sociology and Political Science

• Sociological and political simulations range from public-opinion surveys and election polling to observing everyday behavior and testing social theories.Studies also simulate interview responses and examine population complexity and algorithmic bias.Source: §5.2.2 · pp. 17–18

5.2.3 Online Platform; 5.3 Evaluation

• Online-platform simulations model social information spread and recommendation behavior, while evaluation compares individual fidelity, collective patterns, and system efficiency or cost.Assessments include real-world alignment at micro and macro levels, alongside runtime, resource, scalability, and token or financial costs.Source: §5.2.3 · p. 18, §5.3 · p. 18

6.1 Individual Simulation

• Individual-simulation datasets include descriptions of personal attributes and histories, dialogue data, and evaluation sets pairing profiles with questions or tasks.Dialogue may be single- or multi-turn and can be drawn from plots or social-media utterances.Source: §6.1 · p. 19

6.2 Scenario Simulation

• Scenario-simulation datasets span dialogue-driven and task-driven settings, with QA and multiple-choice formats predominating across easy, medium, and hard tasks.Collection categories include human-annotated, real-world, and synthetic data; domain-specific datasets preserve tailored formats.Source: §6.2 · p. 19

6.3 Social Simulation

• Social-simulation datasets separate agent-initialization profiles and relations from evaluation references describing real-world behavior.Sources include public surveys, existing datasets, and online-platform data.Source: §6.3 · p. 19

7.1 Trend of Individual Simulation

• Individual simulation progresses from coarse models through more nuanced representations to situation-oriented simulations.Early work focused on superficial human traits; later work examined specific individuals, and by May 2024 studies placed them in particular scenarios.Source: §7.1 · p. 19, Figure 5 · p. 22, §7.1.1 · p. 22

7.3 Trend of Society Simulation

• Tables 5 and 6 summarize commonly used datasets for scenario simulation and society simulation, respectively.Source: Table 5 · p. 20, Table 6 · p. 21

7.1.3 Situation-Oriented Simulation

• Situation-oriented simulations move from rule-based game decisions to dynamic responses to environmental feedback and tasks such as dramatic performance, game exploration, and 3D execution.Increasing simulation complexity also raises demands on the underlying architecture.Source: §7.1.3 · pp. 22–23

7.2 Trend of Scenario Simulation

• Scenario simulation advances from simple, single-objective interactions to multi-stage tasks with sequential decisions and task decomposition.Simple scenarios include dialogue and debate frameworks and domain-specific tasks; multi-stage work adds roles and incremental task breakdown, including staged software development.Source: Figure 6 · p. 22, §7.2 · p. 23, §7.2.2 · p. 23

7.2.3 Collaborative Scenario

• Collaborative scenario research emphasizes cooperation and adaptation, including cost-aware social-relation modeling, asynchronous agent trees, and evolving real-world tasks.Applications include embodied environments, resource gathering, medical processes, and social-skills development.Source: §7.2.3 · p. 23

7.3 Trend of Society Simulation

• Society simulation shifts from constructing preliminary environments to scenario-specific alignment and then scaling up with multimodal approaches.Initial environments include an interactive sandbox with agent memories, an epidemic-spread simulation, and infrastructure for adding agents and buildings.Source: §7.3 · pp. 23–24, Figure 7 · p. 23, §7.3.1 · p. 24

7.3.2 Exploring Alignment on Specific Scenarios

• Alignment studies assess whether simulated agents match real situations, using social-network behavior and decisions or macroeconomic phenomena such as inflation and unemployment.The cited work compares simulated social-network behavior at individual and group levels and evaluates macroeconomic agents against rule-based or language-agent baselines.Source: §7.3.2 · p. 24

7.3.3 Scaling Up and towards Multi-Modal

• Scaling work studies large-scale agent interactions and behavioral norms, while multimodal simulations incorporate vision, voice, and other perceptual information.Some work also models perceptual limits and physical demands to support more realistic social interactions.Source: §7.3.3 · p. 24

• The survey categorizes LLM-driven social simulations as individual, scenario, and society simulation, progressing from individual behavior to complex social dynamics.Source: §8 · p. 24
• It systematically reviews architectures, methods, and evaluations across the three categories to provide a structured framework for the field.Source: §8 · p. 24
• The work aims to guide LLM-based simulation development and foster interdisciplinary studies addressing real-world challenges and decision-making.Source: §8 · p. 24