PaperlayerView exact revision on arXiv

Source-linked AI summary

Large Language Models Empowered Agent-based Modeling and Simulation: A Survey and Perspectives

Chen Gao, Xiaochong Lan, Nian Li, Yuan Yuan, Jingtao Ding, Zhilun Zhou, Fengli Xu, Yong Li

arXiv:2312.11970v1cs.AIcs.CLcs.CYcs.MA

TL;DR

Agent-based simulation needs to represent complex, heterogeneous systems, while existing methods face limitations in long-term reasoning, cross-environment generality, and integrative simulation. This survey reviews how LLM-empowered agents address these requirements, organizes recent work across four domains, and identifies open problems and future directions. It concludes that robustness, scaling, open platforms, and ethical risks remain important research directions.

  • Problem

    Existing agent-based simulation methods struggle to combine behavioral accuracy, interpretability, adaptability, and reliability while supporting complex long-term tasks and changing environments.

  • Method

    The paper surveys LLM-empowered agent-based modeling and simulation, analyzes agent requirements and challenges, and organizes applications into cyber, physical, social, and hybrid domains.

  • Results

    The survey comprehensively reviews recent LLM-based agent simulation work and analyzes how LLMs address perception, reasoning, decision-making, adaptivity, and heterogeneity.

  • Takeaways & Limitations

    The survey identifies scaling simulation, open simulation platforms, robustness, and ethical risks as important directions for future research.

  • Takeaways & Limitations

    Current LLM-agent simulations remain vulnerable to robustness problems at both individual and collective levels, including propagating adversarial attacks and exploitable collective behaviors.

Abstract

from arXiv · show

Agent-based modeling and simulation has evolved as a powerful tool for modeling complex systems, offering insights into emergent behaviors and interactions among diverse agents. Integrating large language models into agent-based modeling and simulation presents a promising avenue for enhancing simulation capabilities. This paper surveys the landscape of utilizing large language models in agent-based modeling and simulation, examining their challenges and promising future directions. In this survey, since this is an interdisciplinary field, we first introduce the background of agent-based modeling and simulation and large language model-empowered agents. We then discuss the motivation for applying large language models to agent-based simulation and systematically analyze the challenges in environment perception, human alignment, action generation, and evaluation. Most importantly, we provide a comprehensive overview of the recent works of large language model-empowered agent-based modeling and simulation in multiple scenarios, which can be divided into four domains: cyber, physical, social, and hybrid, covering simulation of both real-world and virtual environments. Finally, since this area is new and quickly evolving, we discuss the open problems and promising future directions.

1 Introduction

The survey addresses the lack of a systematic account of LLM-based agent modeling and simulation by reviewing its foundations, capabilities, application domains, challenges, and future directions.

  • LLM agents support adaptive task performance, human-like responses, self-oriented planning, diverse tool-mediated actions, and interaction with humans or other AI agents.
  • This survey reviews existing large language model-based agent modeling and simulation work and analyzes why LLMs can address agent-based simulation requirements.
  • The survey examines agent capabilities through autonomy, social ability, reactivity, and pro-activeness, alongside perception, reasoning, decision-making, adaptivity, and heterogeneity.
  • It organizes agent-based modeling and simulation applications into physical, cyber, social, and hybrid domains and discusses environment and LLM-driven agent design.
  • The survey identifies future directions including scaling simulation, open simulation platforms, robustness, and ethical risks.

2 Background

Agent-based simulation models heterogeneous agents interacting within environments to explain emergent system dynamics, but traditional approaches face limitations in complexity, generality, planning, and integrative capability.

  • Basic concepts: Agent-based simulation studies macro-level phenomena by modeling heterogeneous agents with attributes, states, adaptive behaviors, and interactions within an environment.
  • Agent capabilities: Agents require autonomy, social ability, reactivity, and pro-activeness to operate independently, interact, respond to environmental changes, and pursue goals.
  • Methodologies: Traditional methodologies include predefined rules, symbolic equations, stochastic modeling, and machine learning models for specifying agent behavior.
  • Limitations: Reactive architectures became prevalent because symbolic deliberative architectures are generally intractable, but direct sense-action loops offer limited deep reasoning and planning.
  • Limitations: Existing approaches struggle to support complex long-term strategies, general agents across diverse environments, and simultaneous descriptive, explanatory, predictive, and hypothetical simulation tasks.

3 Critical abilities of LLM for agent-based modeling and simulation

LLM agents address core agent-based simulation requirements through perception, reasoning, decision-making, adaptation, and richer representation of heterogeneous agents. The paper also identifies persistent challenges in uncertainty, parameter complexity, and capturing the full diversity of human behavior.

  • Perception: LLM agents can perceive and respond to diverse environments through language, supporting the core requirement that agents sense environmental information.Their language capabilities enable direct or indirect comprehension of varied environmental information.
  • Reasoning and decision-making: Traditional rule-based and neural agents may make incorrect or non-optimal decisions because they lack human-level intelligence.Examples include choosing a crowded road or sending an incorrect social-network message.
  • Adaptation: Agent-based simulations must handle uncertainty, uncontrollability, changing environments, and evolving agent states over long-running simulations.The environment and agents may differ substantially from their initial conditions as simulation proceeds.
  • Heterogeneity: Complex parameter settings, limited behavioral data, and difficult validation can force traditional models to oversimplify agent heterogeneity.Parameter selection must account for personal traits, environmental factors, diverse behaviors, and real-world observations.
  • Heterogeneity: Rules or models cannot cover all dimensions of real-world heterogeneity, whereas LLM agents can represent internal cognitive complexity and customized characteristics.Customization can use prompting, in-context learning, or fine-tuning.

4 Challenges and approaches of LLM agent-based modeling and simulation

LLM-based agent simulation requires suitable environments, interfaces, human and domain alignment, personalized roles, action mechanisms, and evaluation strategies. Existing approaches span virtual and real environments and use prompting or tuning to shape agent behavior.

  • Environment and interface: LLM-based agent simulation first constructs a virtual or real environment and specifies how agents interact with it and with one another.The environment must expose information and interaction methods that LLMs can perceive and use.
  • Environment and interface: Existing simulation environments include virtual systems with predefined rules and real-world settings such as economies, websites, and urban environments.Examples include virtual companies and recommender systems, economic activities, web browsing, and city behavior.
  • Environment and interface: Environment interfaces typically represent inputs, outputs, and agent communication through text, with tools extending interaction with complex environments.Tool selection and invocation remain text-mediated even when the underlying environment is richer.
  • Human alignment and personalization: LLM agents may lack domain knowledge and produce irrational decisions, while controllable heterogeneity and personalized role assignment remain difficult.These challenges concern both human-value alignment and adapting agents to distinct simulation requirements.
  • Human alignment and personalization: Prompt engineering aligns outputs by supplying domain instructions, background knowledge, generation patterns, examples, and personalized preferences.Game rules and examples can guide factors such as self-interest and fairness, while preferences can tailor interaction behavior.
  • Human alignment and personalization: Tuning uses domain datasets, expert feedback, or user-specific feedback to fine-tune models toward human knowledge, values, and preferences.Personalization may require selectively constructed datasets, multiple models, or aspect-specific optimization.

4.3 How to simulate actions

LLM agents simulate complex actions by combining planning, memory, and reflection. These mechanisms decompose tasks, preserve experience, and revise behavior using feedback, enabling adaptive strategies across changing simulations.

  • Planning: LLM agents decompose complex goals into manageable subtasks and execute them sequentially to preserve logical progression and contextual relevance.Task decomposition uses the model’s learned patterns and knowledge to structure multi-step problem solving.
  • Planning: Planning methods can dynamically adapt strategies and refine plans in response to ongoing feedback and changing scenarios.The paper presents this as improving decision-making and problem-solving efficiency across contexts.
  • Memory: External memory systems let agents reference experiences and insights that exceed the context window of the underlying LLM.Memory is needed because human-like behavior depends on using large volumes of past information when acting.
  • Memory: Generative memory systems integrate perceptions with continuous experiences and support planning and reflection based on past outcomes.Other systems curate skill libraries and update capabilities through practice and feedback.
  • Memory: Memory also supports adaptation in dynamic social environments by tracking changing social cues, demographics, and prior interactions.In these settings, memory functions as a resource for future social navigation and decision-making.
  • Reflection: Reflection uses evaluators and prompt-based self-critique to refine memory and behavior through continuous feedback.Reflexion combines reward-based assessment with internal performance critique.
  • Integrated action simulation: Planning, memory, and reflection form a cycle in which plans guide action, memory supplies experience, and reflection adjusts strategies.The paper describes this interplay as supporting autonomous decision-making and continuous adaptation.

4.4 Evaluation of LLM agents

Evaluation of LLM-based agent simulations compares outputs with real-world data at individual and population levels, while also examining explanations and ethical risks. Bias, fairness, and harmful outputs require explicit scrutiny alongside simulation accuracy.

  • Realness validation with real human data: Simulation evaluation compares generated outputs with existing real-world data at micro-level and macro-level scales.Micro-level evaluation targets individual behavior or next-state prediction, while macro-level evaluation targets population-level emergent phenomena.
  • Explainability: LLM explanations can be examined to assess whether agents understood simulation rules, perceived environments accurately, and selected actions rationally.This provides an evaluation dimension beyond simulation accuracy.
  • Ethics evaluation: Ethical evaluation should assess bias and fairness across language, culture, gender, race, and other sensitive attributes.The goal is to determine whether generated content perpetuates or mitigates societal biases.
  • Ethics evaluation: Practitioners should inspect generated outputs for harmful or inappropriate content, including hate speech, misinformation, and offensive material.The paper emphasizes this concern because generative outputs are difficult to control.

5 Recent advances of LLM agent-based modeling and simulation

Recent work applies LLM agents to agent-based modeling and simulation across social, physical, cyber, and hybrid domains. Typical applications in the first three domains are illustrated in Figure 2 and detailed in Table 1.

  • Recent advances use LLM agents for agent-based modeling and simulation in social, physical, cyber, and hybrid domains.
  • Typical applications in the social, physical, and cyber domains are illustrated in Figure 2.
  • Details of representative applications are provided in Table 1.

5.1 Social domain I: social sciences

LLM agents are used in social-science simulations to model social dynamics, cooperation, individual behavior, and human responses. The reviewed studies examine behavioral realism, theory-related patterns, collaboration, and the flexibility of agent designs.

  • Social-science simulation: Social-science studies examine whether LLM agents can replicate human behaviors, interactions, and established social laws or patterns.
  • Simulation of social network dynamics: S3 uses real-world social-network data to simulate individual and collective behaviors involving emotions, attitudes, and interactions.
  • Simulation of social network dynamics: Werewolf simulations produce strategic interactions such as trust and confrontation, while transmitted information can reflect human communication biases.
  • Simulation of cooperation: COLA, MAD, and ChatDev organize role-based or debating LLM agents for collaborative problem-solving and decision-making.
  • Simulation of cooperation: Fixed agent roles and collaboration frameworks can lack versatility because their designs are not flexible or adaptable.
  • Simulation of individual social behavior: CoELA combines perception, reasoning, communication, and planning in embodied environments, while Humanoid Agents model needs, emotions, and relational closeness.

5.2 Social Domain II: Economic System

LLM-driven economic simulations span individual decisions, strategic interactions, and market-level systems. Studies report human-like behavior and emergent market dynamics, while also identifying limits in prediction and strategic responses.

  • Simulation taxonomy: Economic simulations are categorized into individual behavior, interactive behavior, and economic system-level simulations.Individual studies examine economic decisions and phenomena; interactive studies focus on games; system-level studies examine markets.
  • Individual economic behavior simulation: LLMs reproduce human-like patterns in behavioral economics experiments, including altruism, fairness preferences, and status quo bias.These experiments include dictator games, fairness constraints, and status quo bias.
  • Individual economic behavior simulation: LLM economic prediction is mixed: stock-market performance can trail state-of-the-art and linear-regression methods, while inflation and finance forecasts can match professional forecasters.The studies suggest basic economic understanding but insufficiently precise perception for accurate prediction without more domain-specific fine-tuning.
  • Interactive economic behavior simulation: Prompted personalities alter repeated-prisoner’s-dilemma cooperation, with competitive and self-interested agents cooperating less than altruistic agents.The same study also reports that agents may respond poorly to opponents’ actions, producing excessive cooperation with betraying opponents.
  • Interactive economic behavior simulation: Game-playing agents can exceed traditional opponents without training, while Cicero outperformed 90% of human players in anonymous online Diplomacy games.These systems extend beyond measuring behavior to constructing stronger strategic agents.
  • Economic system-level simulation: A controlled, scalable, and reproducible framework, Alympics, supports game-theory experiments through a sandbox environment and interacting agent players.The framework separates the environment, which stores and executes game settings, from the agents that interact with it.
  • Economic system-level simulation: Market simulations produce recognizable dynamics, including restaurant popularity divergence, competitor imitation and differentiation, and communication-dependent price collusion.In the Bertrand duopoly simulation, prices approached equilibrium without communication and moved toward monopoly pricing with communication.

5.3 Physical domain

Physical-domain applications of LLM-based agent simulation include mobility, navigation, driving, transportation, and wireless networks. These systems connect language-model agents to spatial or simulated environments for planning and action.

  • Application scope: Physical-domain applications include mobility behaviors, transportation, and wireless networks.The surveyed physical scenarios cover interactions between LLM agents and embodied or infrastructure environments.
  • Mobility and navigation: Research on human mobility examines whether LLMs represent real-world locations and temporal events.Probing studies report spatial and temporal representations in neural layers, alongside investigations of geospatial knowledge.
  • Mobility and navigation: LM-Nav combines large language models with image-language alignment, while LLM-Planner enables few-shot planning for embodied agents.These systems apply LLM capabilities to navigation and embodied-agent planning.
  • Transportation: An LLM driving agent receives car and environmental state from CARLA, selects actions such as stopping or changing lanes, and feeds them back to the simulator.The loop models human-driver behavior through environment observation and action selection.
  • Wireless networks: A multi-agent wireless-network framework supplies human or machine intents through terminals and coordinates task division and planning among on-device LLM agents.The agents interact with the environment and exchange knowledge while solving complex tasks collaboratively.

5.4 Cyber domain

Cyber-domain simulations model behaviors such as information access, website visitation, and online interaction. LLM agents are used to execute web tasks and simulate users within digital platforms.

  • Application scope: Cyber simulations cover information access, website visitation, and network attack or defense behaviors.The domain focuses on human behaviors occurring in cyberspace.
  • Web interaction: WebAgent learns from experience to simulate human behavior on real websites using natural-language instructions.It decomposes instructions, condenses HTML into task-relevant sections, and uses Python programs to interact with websites.
  • Online user simulation: RecAgent models autonomous user behavior and decisions in recommender-system environments.The framework investigates how simulated users may respond to different scenarios on digital platforms.

5.5 Hybrid domain

Hybrid simulations combine domains such as physical movement, social relationships, economics, and geopolitics. The surveyed systems use LLM agents to generate individual, societal, macroeconomic, urban, and international behaviors.

  • Definition: Hybrid simulations jointly consider more than one domain, including physical and social processes.This category covers simulations whose dynamics cannot be confined to a single domain.
  • Social behavior: Generative Agents integrates memory, reflection, and planning to produce believable individual actions and social dynamics.Crowd-sourced evaluations supported these modules, and simulations generated scenarios including parties and mayoral elections.
  • Epidemic simulation: An epidemic simulation links social relationships and perceptions with physical movement and infection susceptibility.Social factors influence epidemic perception, while spatial movement affects exposure-related susceptibility.
  • Geopolitical simulation: War simulations represent countries as LLM agents and use emergent interactions to explain the occurrence of historical conflicts.Evaluated conflicts include World War I, World War II, and the Warring States Period in Ancient China.
  • Macroeconomic simulation: A hybrid macroeconomic simulation scales from tens to hundreds of agents modeling work and consumption in a market.Perception, memory, and action modules support heterogeneity, market understanding, and decisions involving multiple economic factors.
  • Urban simulation: Urban Generative Intelligence combines a digital-twin city environment with CityGPT and LLM-based agents simulating social, economic, mobility, and navigation behaviors.The platform uses city-specific multi-source data and interfaces for embodied agents.

6 Open problems and future directions

Future work centers on making large-scale LLM-agent simulation efficient, broadly accessible, robust, ethically responsible, and better evaluated. The survey also identifies open platforms and alignment methods as important infrastructure and safeguards for this emerging area.

  • 6.1 Efficiency of Scaling Up: Large-scale LLM-agent simulation remains computationally expensive, motivating full-process acceleration to support larger societies and emergent-behavior studies.Batch prompting has achieved up to 5× efficiency improvement in inference token and time costs.
  • 6.2 Evaluation: Benchmarks for LLM-driven agents mainly assess decision and planning, while evaluating agent simulations remains inadequate across behavioral statistics and believability.Existing evaluation approaches examine spatial and temporal distributions or recruit human evaluators, but the survey identifies broader assessment challenges.
  • 6.3 Open Platform: Open platforms could reduce barriers to LLM-driven agent-based simulation and foster community development, but existing platforms mainly target task-solving agents.Examples include Voyager, XAgent, ModelScope-Agent, and BUTLER, which support embodied, real-world, or cross-environment task solving.
  • 6.4 Robustness of LLM-driven ABS: Robustness must be improved against adversarial attacks and out-of-distribution generalization failures at both individual and collective levels.Adversarial effects may propagate among agents, while human-like collective behaviors such as conformity and homophily can create exploitable weaknesses.
  • 6.5 Ethical Risks: Ethical risks include misuse, malicious cyber outputs, medical-agent concerns, and human-like biases that may be amplified through multi-agent transmission chains.Alignment efforts target human instructions, preferences, and values, with Moral Foundation theory used to assess several ethical values.

7 Conclusion

The survey positions LLM-empowered agent-based modeling and simulation as a new approach to constructing intelligent, human-like agents for complex-system modeling. It synthesizes challenges and existing work across four domains, then identifies unresolved limitations and future research directions.

  • The survey examines why LLM agents are needed for agent-based modeling and simulation and how critical challenges can be addressed.
  • Existing LLM-empowered agent-based simulation works are organized into cyber, physical, social, and hybrid domains.
  • For each domain, the survey describes simulation-environment design, LLM-agent construction, and the observations or objectives of simulation.
  • The survey concludes by discussing unresolved limitations and important research directions in this new and rapidly growing area.
Loading 2312.11970v1…