What if LLMs Have Different World Views: Simulating Alien Civilizations with LLM-based Agents

Mingyu Jin, Beichen Wang, Zhaoqian Xue, Suiyuan Zhu, Wenyue Hua, Hua Tang, Kai Mei, Mengnan Du, Yongfeng Zhang

Abstract

This paper presents CosmoAgent, a Multi-Agent System (MAS) utilizing Large Language Models (LLMs) to simulate the evolution and interaction of diverse civilizations. By modeling non-human logic through distinct worldviews—such as Pacifism, Militarism, and Isolationism—and employing mathematical state transition matrices for resource growth, the framework explores complex interstellar scenarios. The study validates historical human growth patterns and investigates emergent behaviors under conditions of asymmetric information and communication delays, providing computational insights into theories like the "Dark Forest" hypothesis.

Introduction

The Search for Extraterrestrial Intelligence (SETI) has long been constrained by the limitations of astronomical observation and the vast scales of cosmic time. While computer simulations offer a potential alternative, traditional models often lack the depth required to capture complex strategic decision-making and diverse ethical frameworks. Furthermore, current LLMs are predominantly Earth-centric, trained on human data and values.

CosmoAgent addresses this gap by using LLMs to model non-human logic and "worldviews." It simulates how civilizations with divergent values interact under conditions of asymmetric information and communication delays, where light-speed limits force agents to act based on outdated observations of their neighbors.

The CosmoAgent Framework

The researchers developed CosmoAgent to integrate mathematical rigor with the flexible reasoning of LLMs. The framework consists of several key components:

Civilization Modeling and Worldviews

Each civilization is defined by a five-dimensional resource vector: Military Capability, Technology Development, Production Capability, Consumption, and Storage. Agents are prompted with specific ideological paradigms to guide their decision-making:

  • Pacifism: Prioritizes mutual benefit and diplomacy.
  • Militarism: Favors pre-emptive strikes to ensure survival in a competitive universe.
  • Isolationism: Focuses on concealment to avoid detection by potentially hostile entities.
Multi-agent architecture of CosmoAgent, featuring distinct civilizational worldviews and the Secretary Agent.

The "Stick" and State Transitions

To model evolution, civilizations apply a 5x5 state transition matrix to their resource vectors. The LLM agent "tunes" this matrix each round, deciding how to allocate resources between growth, defense, and sustainability. This process is managed by the "Stick", a sophisticated archival and modeling tool that records historical trajectories and manages the interplanetary relationship graph.

The 'Stick' framework and interplanetary relationship modeling within CosmoAgent.

Rationality Constraints: The Secretary Agent

To mitigate LLM hallucinations and ensure logical consistency, a Secretary Agent validates all civilizational actions. This agent checks decisions against rationality constraints and sociological theories, preventing impossible scenarios such as instantaneous resource depletion without cause.

Data and Validation

CosmoAgent does not rely on static training datasets. Instead, it employs:

  • Historical Validation: The model's growth curves were validated against historical human GDP data to ensure the mathematical transition matrices align with real-world economic patterns.
  • Synthetic Interaction Logs: The framework generates rich datasets of inter-civilizational "games," capturing strategic maneuvers and communication logs, which have been made available for further research.

Results and Strategic Insights

The simulations yielded several significant findings regarding cosmic sociology:

  • Emergent "Dark Forest" Behavior: Militaristic agents demonstrated systematic targeting of civilizations with the lowest military capability, aiming to eliminate future threats before they could mature. This provides a computational basis for the cautionary warnings of scientists like Stephen Hawking.
  • The Impact of Information Lag: Introducing communication delays significantly altered strategic behavior. Militaristic agents changed their public actions over 60% of the time when forced to rely on outdated information, shifting from immediate strikes to preparatory phases that inadvertently created "windows of opportunity" for targets.
  • Survival Thresholds: Earth’s survival rate at low technological stages is near zero when adjacent to militaristic civilizations. Survival only becomes viable at higher development stages, where the cost of total war becomes prohibitively high for the aggressor.

Discussion and Limitations

While CosmoAgent provides a novel "what-if" engine for cosmic simulation, it faces inherent limitations. The underlying LLMs (GPT-4/3.5) are trained on human data, meaning "alien" logic is still interpreted through a human lens. Additionally, the 5x5 matrix abstraction, while powerful, simplifies the immense complexity of planetary systems. Future work aims to incorporate more diverse ethical paradigms and collaborate with astrobiologists to refine the environmental factors shaping non-human worldviews.

Summary updated

Abstract

This study introduces "CosmoAgent," an innovative artificial intelligence system that utilizes Large Language Models (LLMs) to simulate complex interactions between human and extraterrestrial civilizations. This paper introduces a mathematical model for quantifying the levels of civilization development and further employs a state transition matrix approach to evaluate their trajectories. Through this methodology, our study quantitatively analyzes the growth trajectories of civilizations, providing insights into future decision-making at critical points of growth and saturation. Furthermore, this paper acknowledges the vast diversity of potential living conditions across the universe, which could foster unique cosmologies, ethical codes, and worldviews among different civilizations. Recognizing the Earth-centric bias inherent in current LLM designs, we propose the novel concept of using LLM agents with diverse ethical paradigms and simulating interactions between entities with distinct moral principles. This innovative research not only introduces a novel method for comprehending potential inter-civilizational dynamics but also holds practical value in enabling entities with divergent value systems to strategize, prevent conflicts, and engage in games under conditions of asymmetric information. The accompanying code is available at https://github.com/MingyuJ666/Simulating-Alien-Civilizations-with-LLM-based-Agents.
BibTeX
@misc{jin2024what,
  author = {Mingyu Jin and Beichen Wang and Zhaoqian Xue and Suiyuan Zhu and Wenyue Hua and Hua Tang and Kai Mei and Mengnan Du and Yongfeng Zhang},
  title = {{What if LLMs Have Different World Views: Simulating Alien Civilizations with LLM-based Agents}},
  year = {2024},
  eprint = {2402.13184},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2402.13184},
}
Figure