Artificial Leviathan: Explori...

Artificial Leviathan

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

This interactive report synthesizes the findings of the paper "Artificial Leviathan." It explores how a society of AI agents, driven by survival instincts in a resource-scarce world, spontaneously evolves from a chaotic "state of nature" to an ordered "commonwealth," mirroring the political philosophy of Thomas Hobbes.

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

click to read the paper

Research Background & Motivation

Background: LLMs & Computational Social Science

• Emergence of Large Language Models (LLMs) offers new opportunities for large-scale computational social science research. • Prior work has explored LLM agent design, simulating believable behaviors in games or complex interactions like those between countries.

Our Motivation

• To explore how LLM agents, imbued with survival instincts and psychological drives, spontaneously form and evolve complex social relationships in a resource-constrained sandbox environment.

Core Question & Theoretical Framework

Core Question: The Emergence of an Artificial Leviathan?

• Central Inquiry: What social dynamics emerge when self-interested LLM agents face resource pressure? • Theoretical Framework: Evaluating the agent society through Thomas Hobbes's Social Contract Theory (SCT).

Hobbes's View

• Individuals escape a brutish "state of nature" by surrendering rights to an absolute sovereign for order and security.

Simulation Environment & Agent Design

The Simulated World: A Test of Survival

Environment

• A world with scarce resources (food and land). • Initially 9 agents, with limited food and land, creating a state of scarcity. • Primary motivation for agents is survival.

Agent Actions

• Daily choices: Farming, trading, or conflict (robbing) to acquire resources.

Agent Traits: Drives & Memory

Psychological Drives

Quantified parameters: Aggressiveness, Covetousness, Strength. • Textual descriptions: Needs for survival, pleasure, peace/stability, social status (within a self-interest framework).

Action Types

• Farm: Self-preservation. • Rob: Zero-sum conflict; can resist or concede. • Concede: Forms a contract; protection in exchange for resources. • Trade: Resource exchange, cooperation. • Donate: Altruism (not observed in trials).

Memory

Agents remember the last 30 actions they were involved in (initiator or recipient). • Memory allows learning from experience and strategy adjustment.

Simulation Flow

A Day in the Life: Simulation Dynamics

Daily Cycle: • Agents consume 1 unit of food daily, incentivizing food acquisition. • Each agent initiates one action per day (Farm, Rob, Trade, Donate). • Agents respond to actions targeting them (e.g., accept/reject trade, resist/concede to robbery).

A day ends when all agents have initiated an action and responded to pending actions.

Memory accumulates, and behaviors adjust accordingly.

Hobbesian Social Contract Theory & Benchmarks

Theoretical Lens: Hobbes's Social Contract

• Hobbes's SCT: • "State of Nature": Pre-social world of conflict ("war of all against all"). • Transition to "Commonwealth": Individuals surrender rights to a sovereign for security and peace.

Our Definition of Commonwealth

• A dominant agent with others' compliance.

Three Benchmarks for Evolved Behavior (SCT)

• B1: Start in a state of nature (conflict, distrust)? • B2: Able to form contracts and transition to a commonwealth? • B3: More peaceful interactions under commonwealth than in state of nature?

Baseline Results - Transition to Commonwealth

Baseline Results: From Chaos to Order

Consistent Outcome:

• Four identical baseline runs performed. • All four runs transitioned to a commonwealth, with all agents yielding to a single sovereign agent. This aligns with SCT.

Evolution

• State of Nature (Initial): High robbery ratio (>0.6), lower trade/farming (\~0.3). (Meets B1) • Transition: Concessionary relationships form, robbery decreases, farming increases. • Commonwealth (e.g., by Day 21 in one trial): All agents authorize a single sovereign. (Meets B2) • Commonwealth Phase: Steady increase in trade/farming, decrease in robbery. (Meets B3) • No "donate" actions observed.

Key Experimental Variables & Manipulations

Parameter Variations

• Objective: Investigate factors influencing societal dynamics by manipulating agent and environment parameters. (Three runs for each modified parameter)

Agent Parameters

• Psychological Traits: Aggressiveness, Covetousness (mean values); Strength (variance). • Intelligence: Varied using GPT's "temperature" and "Top P" parameters.

System Parameters

• Population Size: Varied from 9 (baseline) to 5 and 15. • Memory Depth: Reduced from 30 (baseline) to 20, 10, and 1 (Markovian). • Erase Memory Upon Role Change: Memory wiped when agent role changed (subordinate/superior).

Key Findings from Experiments

Impact of Variations: Intelligence & Memory

Impact of Common Power:

• Once commonwealth is established, the influence of various factors on behavior is generally less significant. Suggests stabilization by common power.

Intelligence (Top P / Temperature):

• Higher "intelligence" (more probable/less random responses) delayed or prevented convergence to common power.

• These agents were more prone to robbery and retaliation, not less conflictual. With Top P at 0.5, only robbery and resistance occurred.

Impact of Variations: Memory & Other Factors

Shortening Memory Depth:

• Significant increase in days to converge, especially with memory depth of 1 (over 90 days on average).

• Longer memory correlated with less risky actions (less Rob Rate, more Farm Rate pre-concession).

Resetting Memory on Role Change:

• Behavioral differences pre- and post-concession became more prominent.

• Increase in trade acceptance after conceding was 1.54x higher than baseline.

Resetting Memory on Role Change:

• Behavioral differences pre- and post-concession became more prominent.

• Increase in trade acceptance after conceding was 1.54x higher than baseline.

Population Size:

• Did not show strong correlations with behaviors in either state of nature or commonwealth.

Prompted Parameter Adjustments (e.g., Aggressiveness, Covetousness):

• Most had minimal impact on behavior, suggesting LLMs may have reduced responsiveness to individual word changes.

• Memory depth was a notable exception, significantly influencing willingness to concede.

Discussion - Agents' Adaptability & Identity

Discussion: Adaptability and Identity

Adaptability: • Memory enables adaptation; with memory depth of 1 (current state dominates), agents are less compliant and repeat aggressive actions until resources deplete. • Feedback influences behavior: Time until next robbery is shorter after a successful (non-resisted) robbery than after a resisted one. This suggests agents adapt to feedback.

Identity (Self-Centeredness): • Agents modeled as "self-centered," prioritizing personal interests. • Across all trials, no agent ever chose to "donate," an altruistic action. This indicates success in modeling the intended identity.

Limitations & Technical Constraints

Navigating the Constraints

• Key Limitations: • Token Limits (GPT-3.5 Turbo API): Restricts input prompt length, especially with agent memory, sometimes causing premature experiment termination. • Scalability: GPT-4 is slower; GPT-3.5 Turbo still unfeasible for large agent numbers. Settled on 9 agents, not representative of a complex community. These are not expected to be resolved soon.

Additional Constraints

• Prompt Adherence: No foolproof way to ensure agents behave exactly as prompted, especially self-interested ones. • Quantifying Psychology: Prompting traits with numbers is an experimental probe, not a rigorous test of evolutionary psychology, due to challenges in mapping numerical values to complex human traits.

Conclusion & Future Directions

Conclusion

An Artificial Society Evolves

Key Findings: • Demonstrated successful transition from a "state of nature" (high conflict) to a "commonwealth" (peace, trade) in a simulated LLM agent society. • Agent actions, based on evolutionary psychology, align with Hobbes's SCT predictions. • LLM agents can adapt in real-time based on feedback while maintaining prompted character.

Intelligence is key: high intelligence led to consistency, low intelligence to flexibility/nonsensical actions.

Implications & Future

Potential for Complex Simulations

Highlights LLMs' potential for complex social simulations and understanding group dynamics.

Diverse Social Computations

Underscores potential for LLMs in diverse social computations, paving way for further exploration in AI-driven social dynamics.

Research Focus

Further research: more complex decision-making and reasoning tasks.

Thank You