Do LLM language abilities generalize to graph-structured social networks?

Determine whether large language models (LLMs) generalize to graph-structured objects by establishing whether they can generate social networks that reproduce key structural characteristics of real social networks, such as low edge density and long-tailed degree distributions.

Background

The paper motivates using LLMs for zero-shot, flexible generation of social networks, contrasting them with classical random graph models and deep generative models that either lack realism or require extensive training data. Despite LLMs’ success in simulating human interactions, prior work shows they can struggle with graph reasoning tasks, raising uncertainty about whether their natural language capabilities extend to structured graph objects.

This uncertainty is pivotal for applications like epidemiological modeling and social simulation, where generated networks must plausibly match real-world structural properties (for example, sparsity and heavy-tailed degree distributions). The authors frame this as a central research question driving their paper of prompting methods for LLM-based social network generation.