Synthetic Consumer LabSynthetic Consumer LabTry the demo
Article — 5 min read

LLMs and Generative Agent-Based Models for Complex Systems

The capacity of LLM agents to imitate human behaviour across social systems, and the modelling of complex social dynamics with Generative Agent-Based Models (GABMs).

Research Questions

  1. What transformative role do LLMs play in fields such as network science, evolutionary game theory, social dynamics, and epidemic modelling?
  2. How can the Generative Agent-Based Models (GABMs) framework be used to study complex systems?
  3. To what extent can LLM agents imitate human social behaviour?

Results

  • LLMs can generate human-like behaviours such as fairness, cooperation, and adherence to social norms.
  • Responses can show inconsistency due to prompt sensitivity and underlying model biases.
  • In certain games, LLM agents behaved more cooperatively or more fairly than humans (e.g., Dictator Game, Prisoner’s Dilemma).
  • Multi-agent systems exhibited emergent social dynamics, including homophily and an increased likelihood of repeated interactions.
  • Multi-agent LLM architectures aligned with human behaviour much more closely than single-agent setups (88% vs. 50%).

Findings

  • Human-Like Behaviour:
    • LLM agents displayed behaviour consistent with economic principles such as demand curves and diminishing marginal utility.
  • Inconsistency and Bias:
    • Decisions were influenced even by semantically irrelevant cues such as name or gender.
  • Rationality Differences:
    • Compared with humans, LLMs behaved more fairly in the Dictator Game.
    • They were also more cooperative in the Prisoner’s Dilemma (65% vs. 37% for humans).
  • Context Effects:
    • In epidemic simulations, providing health-related information increased stay-at-home behaviour; supplying community-level statistics further reduced social interactions.
  • Multi-Agent Advantage:
    • In the Ultimatum Game, multi-agent LLM systems captured 88% of human behavioural patterns, outperforming single-agent models.

Scores

  • LLM Models: 5
  • Synthetic Data: 4
  • Method: 5
  • Speed: 2
  • Ethics: 2
  • Accuracy: 4
  • Demographics: 2

If you would like to explore this research in more detail, click here to read the full paper.

Read the full paper ← All articles