Abstract
Traditional information diffusion models often treat messages as static entities, ignoring the nuances of human interpretation and content evolution. This paper presents LAID (LLM-enhanced Agent-based Influence Diffusion), a simulation framework that integrates Large Language Models (LLMs) with Agent-Based Modeling (ABM). By equipping agents with LLM "brains," LAID enables the simulation of information propagation where content is dynamically altered based on agent personalities and social similarities. Our results demonstrate that this approach captures complex "information alteration" phenomena and provides higher interpretability than traditional binary state models.
Introduction
Simulating the spread of ideas, news, or marketing content within social networks is vital for various domains. However, existing "influence diffusion" models, such as the Independent Cascade model, suffer from two primary limitations: content blindness and static propagation. These models typically treat information as a binary state—either a node is "infected" or it is not—without considering the actual textual content or how it might be modified as it passes through the network. In reality, individuals often interpret, summarize, or "spin" information based on their personal biases, a process known as information alteration.
Methodology: The LAID Framework
The LAID framework addresses these limitations by providing each agent with an LLM-driven cognitive component. The system utilizes GPT-3.5-turbo for high-level reasoning and BERT for generating textual embeddings.
Agent Architecture
Each agent is initialized with a unique profile encompassing age, gender, and interests. The core of the agent is its "brain," which processes incoming messages through a multi-step pipeline:
- Sensor: Receives incoming messages.
- Influence Decision: The agent determines whether to be influenced based on content similarity (how well the message aligns with its interests) and social similarity (similarity to the sender).
- Information Alteration: If influenced, the agent uses an LLM to generate a modified version of the message, tailored to its own profile.
- Effector: Sends the newly generated message to its neighbors.
Simulation Parameters
The model introduces a trade-off parameter, $\alpha$, which balances the weight of social influence versus content influence. A high $\alpha$ prioritizes the sender's identity, while a low $\alpha$ emphasizes the message's substance.
Experimental Results
The researchers evaluated LAID across four distinct scenarios: a political election, viral marketing for a smart home hub, the spread of a new idea regarding shared skill spaces, and daily conversation.
Contagion and Alteration
All scenarios successfully exhibited contagion patterns, where information propagated through the network. A key finding was the convergence of the "Alteration Degree." As shown in the figure below, the degree to which messages were modified was highest during the initial stages of propagation and stabilized over time. This suggests that as a message reaches a sub-community of like-minded agents, the pressure for further content modification decreases.
Content Specificity and Creativity
The simulation revealed that "open" content, such as daily conversations, led to higher levels of information alteration compared to specific topics like political elections. Vague prompts allowed the LLM agents more creative freedom to inject their unique personalities into the messages, whereas specific topics constrained the output.
Conclusion and Future Work
LAID represents a significant step toward more realistic social network simulations by moving from probabilistic "dice rolling" to semantic reasoning. While currently limited by the computational cost and scale of LLM API calls, future work aims to scale the simulation to thousands of nodes and incorporate more complex agent behaviors, such as memory and social hierarchies.
Summary updated