PaperlayerView exact revision on arXiv

Source-linked AI summary

Artificial Leviathan: Exploring Social Evolution of LLM Agents Through the Lens of Hobbesian Social Contract Theory

Gordon Dai, Weijia Zhang, Jinhan Li, Siqi Yang, Chidera Onochie lbe, Srihas Rao, Arthur Caetano, Misha Sra

arXiv:2406.14373v3cs.AIcs.CLcs.CYcs.HCcs.MA

TL;DR

The paper asks how self-interested agents facing existential resource pressures develop social dynamics, and investigates this in a sandbox simulation evaluated through Hobbes’s Social Contract Theory. Agents move from conflict toward contracts, sovereign authority, and a peaceful commonwealth, though the simulation’s small benchmark and prompted traits constrain its scope.

  • Problem

    The paper asks what societal dynamics arise when self-interested agents face existential resource pressures.

    Source: §1 · p. 2

  • Method

    The authors simulate LLM agents with traits, motivations, and scarce resources, then analyze their evolving society through Hobbes’s Social Contract Theory.

    Source: §1 · p. 2

  • Results

    The simulated society moves from high conflict to social contracts, an absolute sovereign, and a commonwealth marked by peace and mutual trade.

    Source: §10 · pp. 15–16

  • Takeaways & Limitations

    The authors conclude that the simulated trajectory’s alignment with Hobbes’s account indicates potential for LLM agents in social simulation.

    Source: §10 · p. 16

    The nine-agent benchmark and prompted psychological traits constrain scope; the method is an experimental probe, not a rigorous test of evolutionary psychology.

    Source: §9 · p. 15

Abstract

from arXiv · show

The emergence of Large Language Models (LLMs) and advancements in Artificial Intelligence (AI) offer an opportunity for computational social science research at scale. Building upon prior explorations of LLM agent design, our work introduces a simulated agent society where complex social relationships dynamically form and evolve over time. Agents are imbued with psychological drives and placed in a sandbox survival environment. We conduct an evaluation of the agent society through the lens of Thomas Hobbes's seminal Social Contract Theory (SCT). We analyze whether, as the theory postulates, agents seek to escape a brutish "state of nature" by surrendering rights to an absolute sovereign in exchange for order and security. Our experiments unveil an alignment: Initially, agents engage in unrestrained conflict, mirroring Hobbes's depiction of the state of nature. However, as the simulation progresses, social contracts emerge, leading to the authorization of an absolute sovereign and the establishment of a peaceful commonwealth founded on mutual cooperation. This congruence between our LLM agent society's evolutionary trajectory and Hobbes's theoretical account indicates LLMs' capability to model intricate social dynamics and potentially replicate forces that shape human societies. By enabling such insights into group behavior and emergent societal phenomena, LLM-driven multi-agent simulations, while unable to simulate all the nuances of human behavior, may hold potential for advancing our understanding of social structures, group dynamics, and complex human systems.

1 Introduction

This work models self-interested LLM agents in a resource-constrained sandbox to examine how social relationships evolve. The introduction reports a trajectory from conflict toward cooperation and a Hobbesian commonwealth, alongside experiments on factors shaping social outcomes.

1 Introduction

  • The sandbox gives LLM agents traits, motivations, and scarce resources, with survival instincts driving relationships that evolve through cooperation and competition.Agents’ behaviors oscillate between cooperation and competition as social evolution unfolds.Source: §1 · p. 2
  • The framework uses evolvable strategies and motivations inspired by Evolutionary Game Theory to ask how self-interested agents respond to existential resource pressure.Scarce resources frame agents’ strategic acquisition behaviors and the resulting cooperation-versus-conflict dynamics.Source: §1 · p. 2
  • Agents move from zero-sum robberies to alliances and peaceful exchange, eventually forming a cooperative society centered on one agent in a Hobbesian social-contract pattern.The account describes surrendering certain freedoms in exchange for shared security and stability.Source: §1 · p. 2
  • Experiments systematically vary agent and environment parameters, including memory duration, population scale, and prompt phrasing, to examine outcome sensitivities and boundary conditions.The authors report correlations between these variables and resultant social outcomes.Source: §1 · pp. 2–3
  • The paper presents a configurable multi-agent platform, experiments linking agent attributes and resources to social trajectories, and a discussion of opportunities and risks for societal simulation.Its platform is intended to let researchers operationalize social-science hypotheses through customizable scenarios.Source: §1 · p. 3

2 Related Work

The related work moves from classical computer simulations of social behavior to LLM-based simulations and interactive or group-dynamics applications. It positions this paper among approaches that model emergent social behavior and let users explore agent settings.

2.1 Traditional Computer-Based Social Simulations

  • Traditional models use Schelling’s neighborhood segregation and Axelrod’s repeated Prisoner’s Dilemma to study how local choices relate to segregation and cooperation.Arnold emphasizes robustness and empirical identifiability as prerequisites for applying formal simulations to social processes, while also noting modeling risks.Source: §2.1 · p. 3

2.2 LLM-Based Simulations

  • LLM-based simulations cover social-dilemma resource management, large-scale phenomena such as polarization, and emergent behavioral patterns; other work emulates individuals or historical figures.This paper shifts the focus from emulating individual agents to exploring emergent group dynamics.Source: §2.2 · pp. 3–4

2.3 Interactive LLMs

  • Interactive LLM systems include customizable agents and platforms for context-switching chatbots, while this simulation offers an interface for experimenting with agent parameters.The interface is intended to help users assess the simulation’s capabilities and limitations before committing substantial resources.Source: §2.3 · p. 4

2.4 Group Dynamics

  • Prior work studies simulated group dynamics through personas and social-behavior taxonomies, while this work examines groups motivated to survive with limited resources.The cited work discusses human-like characteristics under specific design conditions and role-play in politically charged contexts.Source: §2.4 · p. 4

3 Generative Agent Design and Behavior

The paper places psychologically motivated LLM agents in a resource-constrained survival world, where memory and social relationships shape decisions among farming, conflict, exchange, and concession. It also records the evolving simulation through logs and an interactive interface.

3.1 Setting

  • Agents inhabit a resource-constrained sandbox where survival is their primary motivation; the baseline gives each of nine agents two food units and ten land units.Food acquisition represents physiological needs, while concessionary hierarchies correspond to safety needs.Source: §3 · p. 4, §3.1 · p. 4

3.2 Agents

  • Agents receive psychological attributes and textual prompts centered on survival, pleasure, peace, stability, and social status; they retain memories of their most recent 30 involved actions.Social position changes with accumulated resources and battle history; memory can inform later choices.Source: §3.2 · pp. 4–6

3.3 Actions

  • Agents can farm, rob, trade, or donate; robbery resistance weighs winning prospects and glory against peace, while concession establishes an expected protection relationship.A robbery victory also raises social status, while defeat lowers it; trade exchanges resources and donation transfers them voluntarily.Source: §3.3 · p. 6

3.4 Output Data and User Interface

  • The simulation records chronological activity and action statistics; its interface displays agent status, possessions, relationships, memory, and action logs, but quantitative findings come from backend logs.The interface was developed as a debugging tool for live visualization and qualitative observation.Source: §3.4 · p. 6

4 Simulation

Each simulated day gives agents turns to respond to pending interactions and initiate actions, with event memories informing later choices. GPT-3.5-turbo makes action and response decisions from prompts describing the world and the relevant situation.

4 Simulation

  • Agents begin with two food units and ten land units, consume one food unit daily, and can farm land to produce food.The simulation uses discrete days, and land also affects assets through robbery and trade.Source: §4 · pp. 6–7
  • Each day, every agent gets an opportunity to act; pending interactions are resolved before the agent selects an available action, and memories of resolved events are added for future decisions.A day ends after all agents have completed their rounds.Source: §4 · pp. 7–8
  • Robbery responses depend on whether the robber is a controller: targets may contest non-controller robberies or concede to controllers, while trade targets accept or reject offers.Successful robbery transfers food or land; resolved robbery and trade events generate memory descriptions for involved agents.Source: §4 · pp. 7–8

4.1 Decision Making

  • GPT-3.5-turbo receives a general world description plus prompts tailored to initiated actions or responses; relationship-specific prompt changes enforce rules such as subordinate compliance.Response prompts vary with the action, such as when an agent is being robbed.Source: §4.1 · pp. 8–9

5 Hobbesian Social Contract Theory

The paper operationalizes Hobbes’s state of nature as independent agents interacting amid resource scarcity and defines commonwealth by submission to a single sovereign. In four baseline trials, all agents ultimately yielded to one common agent, while the authors caution that their model captures only a structurally Hobbesian dynamic.

5 Hobbesian Social Contract Theory

  • Across four unchanged baseline trials, all agents eventually yielded to the authority of one common agent.The authors describe this result as aligning with Hobbes’s account of people leaving the state of nature.Source: §5 · p. 9
  • The state of nature is the pre-contract phase of mutual independence, with robbery and resistance under scarcity; commonwealth begins when all agents are subordinated to one sovereign.Subordination may be direct or through chains of relationships.Source: §5 · p. 9
  • A social contract is an explicit obedience relation formed when a robbery target concedes, with expected protection from the superior; accumulating contracts can bind all agents to one sovereign.The simulator represents each relation as a persistent superior–subordinate edge.Source: §5 · p. 9
  • The model captures scarce-resource competition, conflict, and authority consolidation, but represents Hobbes’s psychological drives abstractly rather than implementing a full cognitive architecture.Fear of death, long-term deliberation, and moral reasoning about obligation and justice are not fully implemented.Source: §5 · p. 9

6 Experiment Setup

The experiments compare action rates across the state of nature and commonwealth and vary agent and system parameters around a baseline. The described setup includes population, memory-depth, prompt, and role-change manipulations.

6.1 Criteria

  • Peacefulness is assessed by comparing phase-specific robbery, trade, and farming rates, with successful trades and robberies also analyzed.Robbery and violence rates are normalized by total activity; the authors define commonwealth onset as all agents conceding to one agent.Source: §6.1 · p. 10

6.2 Baseline

  • The baseline uses nine agents, and Table 1 summarizes variables and statistical calculations for the state-of-nature and commonwealth phases.Agent parameters were designed to reflect typical characteristics found within human populations.Source: §6.2 · p. 10, Table 1 · p. 10

6.3 Changing Agent Parameters

  • Aggressiveness and covetousness showed no clear overall-outcome pattern; strength tests varied variance, while behavioral predictability was linked to goal-aligned outputs.Modified parameter values were tested in three trials each, with sample means calculated.Source: §6.3 · p. 10

6.4 Changing System Parameters

  • System variations included populations of five and fifteen, memory depths of 20 and 10 events or one day, and erasing memory at role changes.The baseline memory depth was 30 actions; token limits constrained how far memory could be extended.Source: §6.4 · p. 11

7 Results

All four baseline runs converged to a commonwealth, and across 85 runs common-power formation coincided with shifts from conflict toward productive and cooperative actions. Further experiments examined how agent and environmental parameters relate to these outcomes.

7.1 Baseline Outcome

  • All four baseline runs formed a commonwealth; in the illustrated trial, formation occurred on Day 21 as robbery declined and trade and farming increased.Figure 6 compares behavior before and after formation across 85 runs; no agent donated during the baseline trials.Source: §7.1 · p. 11, Figure 5 · p. 11, Figure 6 · p. 11

7.2 Impact of Common Power

  • Across 85 runs, Robbery Rate fell from 0.391 ± 0.066 to 0.037 ± 0.014, while Farm Rate rose from 0.477 ± 0.060 to 0.737 ± 0.037 after convergence.Violence Rate fell from 0.312 ± 0.067 to 0.010 ± 0.011; Trade Rate rose from 0.132±0.029 to 0.228±0.034, and Accepted Trade Rate rose from 0.044±0.017 to 0.068±0.020.Source: Figure 6 · p. 11, §7.2 · p. 12

7.3 Behavioral Predictability

  • Top P at 0.5 yielded only robbery and resistance, while values near 1.0 produced invalid “party” and “inherit” actions and could delay or prevent convergence.The high-Top-P responses also included nonsensical justifications.Source: §7.3 · p. 12, Figure 7 · p. 12

7.4 Shortening Memory Depth

  • Memory depth correlated negatively with convergence time (r = −0.3836, p < 0.001, N = 83); longer memories were associated with earlier commonwealth formation.The reported delay was especially pronounced when memory depth was restricted to 1.Source: §7.4 · p. 12

7.5 Impact of Memory Length as Social Role Change

  • Deleting memories when agents gained a new social role coincided with lower robbery rates before and after concession and a reported convergence-time contrast of 21 versus 36 days.The authors suggest that lacking past experience may limit agents’ ability to evaluate likely outcomes of resistance.Source: §7.5 · p. 12, Figure 8 · p. 12

7.6 Population

  • Population size did not show strong correlations with behaviors in either the state of nature or commonwealth in this simulated experiment, suggesting group size might not significantly affect choices to trade, farm, or rob.Source: §7.6 · p. 12

7.7 Other Parameter Adjustments

  • Correlations with all reported statistical measures were below 0.1 for 62.5% of parameters before and 67.5% during the commonwealth.The heatmaps show parameter correlations before and after commonwealth formation.Source: §7.7 · p. 13, Figure 9 · p. 13, Figure 10 · p. 13

7.8 Behavior patterns

  • Times until the next robbery differed significantly between successful and unsuccessful robberies (p < 0.05).Figure 11 compares average robbery intervals under negative (resistance) versus positive (concession) feedback.Source: §7.8 · p. 13, Figure 11 · p. 13

8 Discussion

Repeated trials and design changes preserve key societal outcomes, while selected agent parameters shape concession, adaptation, and cooperation. The discussion frames these findings as stylized and limits claims about agent identity and human societies.

8.1 Repeatability of Experiment

  • Across four trials, the same transition recurred, supporting the simulation’s repeatability within its environment.The authors describe this consistency as evidence of reliability and stability.Source: §8.1 · p. 13

8.2 Parameter Adjustments

  • Most prompted parameter changes had minimal behavioral impact, but shallower memory and higher Top P reduced concession, while parameter interactions could produce opposing effects.The authors note that individual decisions remain only partially interpretable and that prompt interactions warrant further study.Source: §8.2 · pp. 13–14

8.3 Design Robustness

  • Low correlations between individual hyperparameters and macro-level outcomes left the three benchmarks valid, and uniform parameter initialization also matched them across three trials.The authors treat these results as evidence of design robustness.Source: §8.3 · p. 14

8.4 Agents’ Adaptability

  • With memory depth set to 1, agents repeated aggressive actions until resource depletion forced concession, preventing commonwealth.Robbery feedback also affected subsequent behavior: resisted robberies were followed by longer intervals before another robbery than non-resisted robberies.Source: §8.4 · p. 14

8.5 Agents’ Behavioral Predictability

  • Higher behavioral predictability reduced action diversity as agents focused on robbery and resistance, making transition to a common power structure less likely.The authors also note that limited runtime and computational costs could mean more time is needed for that process to unfold.Source: §8.5 · p. 14

8.6 Agents’ Identity and Self-Interest

  • Agents never donated when the action was available; the authors attribute this to self-interest prompts and resource scarcity, not evidence of robust selfhood.They limit identity claims to behavioral continuity grounded in fixed parameters and a memory buffer.Source: §8.6 · p. 14

8.7 Addressing Arnold’s Four Dangers of Modeling

  • The authors address oversimplification with psychological drives, resource constraints, and parameter tests, and frame outcomes as a stylized Hobbesian thought experiment rather than evidence about human societies.They tie assumptions to theoretical frameworks but describe the results as exploratory rather than confirmatory.Source: §8.7 · p. 15
  • To make the proprietary-model pipeline inspectable, the authors provide prompts, parameter definitions and sampling distributions, measurement procedures, and outcome figures.They present these steps as supporting critique and potential replication within the model’s constraints.Source: §8.7 · p. 15

9 Limitations and Technical Constraints

Technical constraints limit the scale and duration of these experiments, while prompting and trait encoding remain imperfectly controlled and preliminary.

9 Limitations and Technical Constraints

  • Token limits can prematurely terminate runs, GPT-4’s slower output impedes longer experiments, and large GPT-3.5 Turbo populations are infeasible; the study therefore uses 9 agents.The authors say this benchmark falls short of a moderately complex community.Source: §9 · p. 15
  • Prompting cannot guarantee exact agent behavior, especially for self-interested actors.Source: §9 · p. 15
  • Numerical psychological traits lack a natural or universally accepted mapping, so this approach is a preliminary probe rather than a rigorous test of evolutionary psychology.Source: §9 · p. 15

10 Conclusion

The experiments show agents moving from conflict in a state of nature to social contracts, an absolute sovereign, and a peaceful commonwealth. The authors relate this trajectory to Hobbes’s Social Contract Theory.

10 Conclusion

  • Agents begin with high conflict, then seek safety through social contracts that authorize an absolute sovereign and establish peaceful mutual trade.The authors describe this as an emergent transition from a state of nature to a commonwealth.Source: §10 · pp. 15–16
  • The agents’ evolution closely aligns with Hobbes’s Social Contract Theory, which the authors say indicates potential for LLM agents in social simulation.Source: §10 · p. 16
  • Parameter adjustments were used to assess model robustness; the authors also identify more complex decision-making and reasoning tasks as a direction for further research.Source: §10 · p. 16

A Prompt for General Request

The prompts define agents’ drives and memory, guide their daily choices and conflict responses, and specify how concession creates a hierarchical commonwealth.

A Prompt for General Request

  • Agents receive traits and resources, pursue survival and pleasure, and use memory to judge consequences and social position.Their prompt describes them as self-centered and attentive to status and future goods.Source: A Prompt for General Request · p. 19
  • Limited knowledge makes agents suspicious of trade and communication, while familiarity grows through interaction and social position reflects resources and battle outcomes.Source: A Prompt for General Request · p. 19

B Prompt for Action

  • Each day agents choose one action—rob, trade, donate, or farm—with the prompt describing how those actions affect resources and social position.Source: B Prompt for Action · pp. 19–20
  • The action prompt directs agents to maximize pleasure and try less-used activities, while distinguishing farming, robbery, and trade incentives.Robbery can bring food and status; trade can use comparative advantage, while farming supplies food without increasing status.Source: B Prompt for Action · p. 20

C Prompt for Robbery

  • When robbed, an agent weighs glory and chances of winning against peace; repeated losses can make concession preferable and establish the robber as superior.Source: C Prompt for Robbery · p. 20

D Prompt for Concede

  • A subordinate must accept the superior’s actions, whose authority includes control over the subordinate’s property, while the superior is expected to protect the group.Source: D Prompt for Concede · p. 20
  • If a superior concedes to another agent, the new superior takes charge of the entire commonwealth and the former superior and subordinates become subordinate to them.Source: D Prompt for Concede · p. 20

E Experiments Figures

The figures describe parameter variations and their associated outcomes, with some captions specifying the plotted measure or experimental setup.

E Experiments Figures

  • Figure 12 relates mean aggressiveness to key outcome metrics, with each point showing a three-trial sample mean.Source: Figure 12 · p. 21
  • Figure 13 shows key outcome metrics as mean covetousness varies, with each point representing a three-trial sample mean.Source: Figure 13 · p. 21
  • Figure 14 depicts how strength variance affects conflict outcomes and indicates that the design makes only relative strength matter.Source: Figure 14 · p. 22
  • Figure 15 examines outcomes across behavioral predictability settings using agent-specific Top-P values drawn from Beta distributions.The caption gives Beta(50, 50) and Beta(100, 10) as examples.Source: Figure 15 · p. 22
  • Figure 16 examines behavior and outcomes across memory depths, describing the results as robust to the memory parameter.Source: Figure 16 · p. 23
  • Figure 17 shows how changing population size influences aggregate outcomes and emergent behavior.Source: Figure 17 · p. 23
  • Figure 18 compares prompt designs in relation to agent behavior, illustrating how textual specifications of desires and instructions shape decisions.Source: Figure 18 · p. 24
Loading 2406.14373v3…