Reproduce “Homophily-induced Emergence of Biased Structures in LLM-based Multi-Agent AI Systems” on OpenResearch

About this paper

Aliakbar Mehdizadeh, Martin Hilbert

This research investigates how large language model (LLM)-driven agents form networks, demonstrating that incorporating social attributes leads to the emergence of homophilous and structurally biased network structures. It reveals that LLMs embed human-like preferences, with value-based attributes like political orientation causing stronger fragmentation than status-based attributes, and exhibits model-specific differences in generating these biases.

Key insights

• LLM agents, when provided with social attributes, embed and amplify human-like homophilic preferences, generating socially biased network structures that deviate significantly from purely structural attachment.
• "Value homophily" attributes (like political orientation and religious practice) induce stronger network fragmentation and polarization than "status homophily" attributes (such as age, education, and gender).
• Significant differences exist between LLMs in how they process attribute information, leading to varying levels of homophily, disassortativity, and overall network compactness, acting as "socio-cognitive mirrors" of biases.

Summary by alphaXiv.