Integrating LLM and Diffusion-Based Agents for Social Simulation

Research introduces HySID, a hybrid framework designed to improve individual-level information adoption prediction by integrating Large Language Models (LLMs) with conventional information diffusion models. This approach addresses the computational cost and unreliability of LLMs for large social networks with sparse user histories, while enhancing traditional diffusion models' limited understanding of semantic content and user-item compatibility. HySID achieves this by adaptively selecting structurally informative core users from historical data for more efficient and accurate simulation.

Why it matters

This development is strategically important as it addresses critical limitations in predictive analytics for information spread within complex networks. By enhancing the efficiency and accuracy of user behavior prediction, organizations can better anticipate trends, manage information flow, and develop more targeted strategies across various operational domains.

Key insights

• Large Language Models (LLMs) offer robust semantic reasoning for user modeling but are computationally intensive and unreliable for social simulations with sparse user behavioral histories.
• Conventional information diffusion models are efficient at leveraging historical propagation patterns and social structures but lack the capability to understand item content and user-item semantic compatibility.
• HySID, a new hybrid framework, aims to combine the strengths of LLMs (semantic reasoning) and diffusion models (structural efficiency) for individual-level information adoption prediction.
• The framework improves efficiency by adaptively selecting a small set of structurally informative 'core users' from historical user-relation graphs, reducing the need for computationally expensive LLM applications across an entire network.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2510.16366