LLM-Based Multi-Agent Systems are Scalable Graph Generative Models

TL;DR

This work presents GraphAgent-Generator (GAG), an LLM-based, simulation-driven framework for large-scale, text-attributed social graph generation. By casting actors and items as LLM agents within a bipartite graph and evolving the network over multiple rounds via the S-RAG interaction mechanism and parallel acceleration, GAG achieves realistic macro- and micro-scale network properties, including power-law degree distributions and small-world topology, while preserving textual feature correlations. The framework scales to roughly 100,000 nodes or 10 million edges and delivers substantial speed-ups, enabling efficient exploration of graph growth in diverse domains. The results indicate that agent-driven simulations can produce high-quality, interpretable graphs that align with real-world network characteristics, with open-source code provided for replication and extension.

Abstract

The structural properties of naturally arising social graphs are extensively studied to understand their evolution. Prior approaches for modeling network dynamics typically rely on rule-based models, which lack realism and generalizability, or deep learning-based models, which require large-scale training datasets. Social graphs, as abstract graph representations of entity-wise interactions, present an opportunity to explore network evolution mechanisms through realistic simulations of human-item interactions. Leveraging the pre-trained social consensus knowledge embedded in large language models (LLMs), we present GraphAgent-Generator (GAG), a novel simulation-based framework for dynamic, text-attributed social graph generation. GAG simulates the temporal node and edge generation processes for zero-shot social graph generation. The resulting graphs exhibit adherence to seven key macroscopic network properties, achieving an 11% improvement in microscopic graph structure metrics. Through the node classification benchmarking task, we validate GAG effectively captures the intricate text-structure correlations in graph generation. Furthermore, GAG supports generating graphs with up to nearly 100,000 nodes or 10 million edges through large-scale LLM-based agent simulation with parallel acceleration, achieving a minimum speed-up of 90.4%. The source code is available at https://github.com/Ji-Cather/GraphAgent.

Figures (11)

• Figure 1: An illustration of the GAG Framework for generating social graphs: (1) Node Formulation, where actor and item node sets are initialized; actor nodes are initilized with LLM-based agents, while item nodes are managed by an item agent; (2) Interaction Simulation, where agents engage in pair-wise interactions within a virtual environment; (3) Graph Projection, where the actor-item bipartite graph is folded along specified node or edge types.
• Figure 2: The power-law Distribution of degrees in generated graphs. The degree $k$ is plotted against the probability density function $P_k$ on a log-log scale, where $\alpha$ denotes the exponent parameter, $k_{min}$ represents the cut-off $k$alstott2014powerlaw.
• Figure 3: The Shrinking Diameter phenomenon simulated by GAG; The left figure demonstrates that as the graph evolves, $D_{e}$ gradually decreases in action and follows networks; The right figure presents an ablation experiment of the ReRanking, demonstrating its effect on $D_e$ in friend network.
• Figure 4: Periodic Variation of Degree in Movie Rating Network; Figure \ref{['fig:movielens_p']} shows the number of released movies over time and the degree of the movie rating network over time in MovieLens dataset; Figure \ref{['fig:movie_rating_p']} also shows the number of released movies and the degree of the movie rating network over time in GAG.
• Figure 5: The proportion of the largest connected component grows steadily over time.