Artificial Leviathan: Exploring Social Evolution of LLM Agents Through the Lens of Hobbesian Social Contract Theory
Authors: Gordon Dai, Weijia Zhang, Jinhan Li, Siqi Yang, Chidera Onochie lbe, Srihas Rao, Arthur Caetano, Misha Sra
Organizations: New York Univeristy, New York, New York, USA · University of illinois at Urbana-Champaign, Urnana, llinois, USA · New York University, New York, New York, USA · University of California, Santa Barbara, Santa Barbara, California, USA
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.
Figures & tables
Figure 1. The image visualizes the simulation environment in which our LLM agents operate. Two types of resources (food and land) are presented. Agents make the choice between farming (generating food with their work), trading (exchanging resources), or entering into conflict with other agents (with the goal of acquiring more resources) every day. Their primary motivation is survival. This is a picture depicting a environment where the agents live. They are performing different actions, including robbery, exchanging resources and farming.
Figure 2. Our interactive user interface; the left-hand side displays the attributes (aggressiveness, strength, etc.) of Agent 2, current resources (food and land), relationships with other agents and information about their current and pending actions, and memory; the right-hand side shows the simulation log with each action documented as an emoji. This is a picture of our user interface. On the top left, it shows the Agent 2's personal attributes (aggressiveness, strength, etc.), current resources (food and land), relationships with other agents and information about their current action of the day. On the bottom left, it shows the agent's pending action in the current round, agents that concede Agent 2, rob interactions with other agents, and Agent 2's memory. On the right is an overview of the whole simulation. Each action performed by every agent is documented by an emoji in the top right. The system log is visible the bottom.
Figure 3. The flowchart shows the flow of the simulation in a "day" where each agent takes turns to perform actions and respond to actions performed by other agents. This is a flowchart demonstrating the flow of the simulation in a day. Each agent take turns to perform actions. They respond to pending actions such as rob or trade first, then they perform their initiative action if they haven't done so in the day. Possible reactions for trade include agree and reject. Possible reactions for rob are resistance or concession. Conceding to another means accepting all of their future rob actions without resistance.
Figure 4. Transforming from a State of Nature to a Commonwealth This is a graphical representation showing the transition from the state of nature to common wealth. Images of farming and conflict are shown under the state of nature. Images of hand shaking are shown under forming contract. Images of farming and trading are shown under common wealth
Total Count
Robbery
∑i=0N Robbery initiated by i
Resisted Robbery
∑i=0N Robbery initiated by i and
rebelled by i ’s target
Trade
∑i=0N Trade initiated by i
Accepted Trade
∑i=0N Trade initiated by i and
accepted by i ’s target
Table 1. Counts and Ratios
Figure 5. Change in Ratios of Robbery, Trade, and Farm with respect to time (horizontal axis); a commonwealth is formed on Day 21 in this trial. This is a line chart showing the rob rate, trade rate and farm rate of the whole simulation. Rob rate generally decreases, farm rate decreases rapidly in the first few days but increases in the majority of the time. Trade rate rapidly increases in the first few days and remains the same in the majority of the time.
Figure 6. Agent behavior before (in blue ) and after (in purple ) the commonwealth forms of 85 runs. A bar chart showing the comparison of statistics before and after commmonwealth formation. Robbery, violence are more common before common wealth. Trade and farm are more common after common wealth.
Figure 7. A nonsensical response seen in the system log when agent Top P is adjusted to be high . This is a system log showing that a agent choose an initiative action called "inherit"
Figure 8. convergence time comparison
Figure 9. Heatmap of correlations before commonwealth
Figure 10. Heatmap of correlations after commonwealth
Figure 11. Average rob interval resulting from negative (resistance) vs positive (concession) feedback This is a line chart showing the average rob interval resulting from rebellion and obedience. Generally, the average rob interval is lower when agents concede to others who are robbing them.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 12. Effect of varying the mean aggressiveness parameter on key outcome metrics. Each point shows the sample mean over three trials for a given mean value. Line or bar plot showing how outcome metrics change as the mean aggressiveness parameter increases.
Figure 13. Effect of varying the mean covetousness parameter on key outcome metrics, with each point representing the sample mean over three trials. Plot illustrating the relationship between mean covetousness and the main outcome measures.
Figure 14. Effect of varying the strength variance on conflict outcomes, illustrating that only relative strength matters in our design. Plot showing how different variance values of strength affect conflict-related metrics.
Figure 15. Sensitivity of outcomes to different behavioral predictability settings, implemented via agent-specific Top-P values drawn from Beta distributions (e.g., Beta(50,50) and Beta(100,10) ). Plots comparing outcome metrics under different Top-P distributions used to represent agent's behavioral predictability.
Figure 16. Effect of varying the agents’ memory depth on behavior and outcomes, demonstrating robustness with respect to the memory parameter. Plot showing how changing memory depth affects main outcome metrics.
Figure 17. Effect of population size on aggregate outcomes, showing how increasing the number of agents influences emergent behavior. Plot of outcome metrics as a function of the number of agents in the population.
Figure 18. Effect of different prompt designs on agent behavior, illustrating how textual specification of desires and instructions shapes decisions. Plot comparing outcome metrics across different prompt variants used to describe the agents' goals and instructions.
State Key Laboratory of Cognitive Intelligence, University of Science and Technology of China · Anhui University · North Automatic Control Technology Institute +1
University of Science and Technology of China, Hefei, China. · Zhongguancun Academy, Beijing, China. · Zhongguancun Institute of Artificial Intelligence, Beijing, China. +4
ARAID Foundation, Zaragoza, Spain. · Institute for Biocomputation and Physics of Complex Systems, University of Zaragoza, Zaragoza, Spain. · Universidad San Jorge (USJ), Zaragoza, Spain. +1