Generative Social Network Simulation

Motivation

Standard ABM (agent-based modeling) uses hand-coded behavioral rules. LLMs offer a different possibility: agents that reason about their social context and update beliefs through natural language. This project asks whether LLM-driven agent simulation can model realistic opinion dynamics in a graduate student social network.

Design

I defined six graduate student archetypes — the Pragmatist, the Theorist, the Networker, the Burnout, the Overachiever, and the Skeptic — each with a system prompt encoding their epistemic tendencies, social preferences, and prior beliefs on academic/career topics.

Agents are placed on a synthetic Watts-Strogatz small-world graph. At each timestep, an agent receives a "message" from a neighbor (a sampled opinion), generates a response (GPT-4 call), and updates their stated belief. Opinion is measured as a scalar extracted from the response via a secondary LLM call.

Evaluation

• Polarization index: variance of opinions across agent population over time
• Echo chamber detection: within-cluster vs. between-cluster opinion similarity
• Belief change rate by archetype under heterogeneous information exposure

Findings

LLM agents exhibit emergent opinion polarization without explicit polarization reward — consistent with empirical social network dynamics. The Pragmatist archetype shows highest belief plasticity; the Theorist shows lowest. Network homophily amplifies polarization as expected.

This work connects to my essay You Are Not My Type!

Settled · 2025–26