Learning dynamics from online-offline systems of LLM agents

1. Problem Statement

The paper addresses the growing concern regarding the interplay between online and offline systems, specifically how automated Large Language Model (LLM) agents influence the spread of information and social instability. While traditional research has focused on human behavior or simple bot activity, the complexity of LLM-driven agents—whose behaviors are defined by prompt-engineered "personalities"—remains under-studied from a mathematical modeling perspective.

The core research questions are:

• How do different types of offline events (e.g., peaceful protests vs. violent conflicts) propagate through a network of LLM agents?
• How do assigned personality traits (based on the Big Five model) affect the transmission rates of information?
• Can the complex, stochastic dynamics of LLM interactions be accurately captured by simplified mathematical models (specifically mean-field approximations)?

2. Methodology

A. Simulation Setup

The authors constructed a multi-agent simulation environment to model information spread:

• Network Structure: A $k$-regular random graph with $N=128$ agents and degree $k=5$.
• Agent Profiles: Each agent is assigned a unique "personality" based on the Big Five Model (Openness, Conscientiousness, Extraversion, Neuroticism, Agreeableness). Each trait is binary (High/Low), resulting in $2^5 = 32$ distinct personality profiles. These are distributed uniformly across the network.
• Event Data: News events were sourced from the ACLED (Armed Conflict Location & Event Data) database, focusing on the Philippines (2021–2024).
  • Events were clustered into two groups using Principal Component Analysis (PCA) on text embeddings: Cluster 1 (Severe/Violent) and Cluster 2 (Peaceful/Demonstrations).
• Simulation Process:
  1. A "seed" event is introduced to one agent.
  2. Agents decide whether to share the news with neighbors based on their personality and the event content, using GPT-3.5-turbo-1106.
  3. The decision process is stochastic (temperature set between 0.7 and 1.3) to mimic human-like noise and creativity.
  4. The simulation runs for $T=15$ iterations, tracking the ratio of "engaged" (shared) vs. "unengaged" agents.
  5. The experiment was repeated 5 times for 10 different seed events (5 per cluster).

B. Mathematical Models

The authors proposed and fitted two distinct models to the simulation data:

1. Stochastic Agent-Based Network Model (Logistic Regression):

   • Models the probability $p_i$ of an unengaged node becoming engaged.
   • Uses a logistic function $\sigma(z)$ where the input is a linear combination of the 32 personality profile weights ($\beta_t$) and the event severity ($\beta_s$).
   • Parameters are estimated via Maximum Likelihood Estimation (MLE).

2. Mean-Field Differential Equation Model:

   • A continuous approximation of the discrete agent-based process, analogous to Susceptible-Infected (SI) epidemiological models.
   • The population is divided into two groups based on behavioral clustering (Group 1: Low transmission; Group 2: High transmission).
   • The system is governed by coupled ODEs:
     $$\frac{dU_i}{dt} = -k\tau_i U_i E$$
     $$\frac{dE}{dt} = \sum_{i=1,2} k\tau_i U_i E$$
     Where $U_i$ is the unengaged fraction of group $i$, $E$ is the engaged fraction, $k$ is the degree, and $\tau_i$ is the transmission rate for group $i$.
   • Parameters ($\tau_1, \tau_2$) are estimated by minimizing the Least Squares Error between the model curve and the observed simulation data.

3. Key Contributions

• Framework for LLM Dynamics: Established a rigorous methodology for simulating and mathematically modeling information cascades in networks of LLM agents with distinct, prompt-defined personalities.
• Model Comparison: Demonstrated that a low-dimensional mean-field differential equation model (with only 2 parameters) outperforms a complex, high-dimensional agent-based logistic model (with 33 parameters) in predicting global engagement dynamics.
• Personality Clustering: Identified that despite 32 unique profiles, LLM agents naturally cluster into two distinct behavioral groups regarding information sharing: a "low-transmission" group and a "high-transmission" group.
• Event Severity Impact: Quantified how event severity (peaceful vs. violent) modulates sharing probabilities, finding that peaceful events generally spread more easily than violent ones in this specific setup.

4. Results

Model Performance

• Mean-Field Model Superiority: The 2-parameter mean-field model achieved a Root Mean Squared Error (RMSE) of 0.0223 against the simulation data. In contrast, the full 33-parameter logistic model had an RMSE of 0.0414.
• Reasoning: The mean-field model's superior performance is attributed to its parameter estimation method (minimizing curve error directly) versus the logistic model's MLE approach (minimizing individual decision likelihood), which may not perfectly align with aggregate dynamics.
• Event Type Fit: Both models fit peaceful events better than severe events. The mean-field model achieved an RMSE of 0.0098 for peaceful events, nearly matching the observed behavior exactly.

Behavioral Insights

• Personality Drivers: Statistical analysis revealed that Openness and Extraversion are the strongest predictors of information sharing.
  • Openness: A sharp, non-linear jump in sharing probability occurs when Openness shifts from Low to High.
  • Extraversion: Positively correlated with sharing.
  • Conscientiousness: Showed a negative (though less significant) correlation.
• Transmission Rates: The inferred transmission rate for the high-activity group ($\tau_2 \approx 0.285$) was significantly higher than the low-activity group ($\tau_1 \approx 0.013$).
• Dynamics: The global engagement curve is driven initially by the rapid activation of Group 2, followed by a slower, gradual increase as Group 1 agents eventually engage.

5. Significance and Future Directions

• Predictability of AI Swarms: The study suggests that despite the complexity of individual LLM behaviors, large-scale LLM networks exhibit stable, predictable spreading patterns that can be modeled with simple differential equations.
• Policy and Safety: Understanding these dynamics is crucial for anticipating how AI agents might amplify misinformation or stabilize social unrest in real-world online-offline feedback loops.
• Future Work: The authors propose extending this framework to Human-LLM interaction networks, where LLMs interact with real human users. This could serve as a controllable testbed for studying complex social phenomena and collective behavior at scale.

In conclusion, the paper successfully bridges the gap between complex agent-based simulations and tractable mathematical theory, proving that mean-field approximations are highly effective tools for understanding and forecasting information propagation in LLM-driven social systems.