Simulated Social Behavior: How LLM Agents Can Reshape Behavioral Economics

Behavioral economics looks into the “why” behind our economic choices – why we cooperate, compete, or take financial risks, and how social norms and emotions shape our decisions. Traditionally, these behaviors are studied through human experiments, which, while insightful, are complex, costly, and often ethically challenging.

A recent study by researchers from the University of Tokyo, on “Spontaneous Emergence of Agent Individuality Through Social Interactions in LLM-Based Communities”, introduces a transformative approach to this field. By using LLM-based agents, the researchers demonstrated that complex social behaviors, emotional responses, and even individual personalities can emerge within a virtual society without any preassigned characteristics.

In this blog post we will be showcasing how and why we can leverage agents for observing economic and social behaviors in a controlled, flexible, and scalable environment, highlighting an exciting opportunity for behavioral economics.

How It Works: The Technical Backbone of LLM Agent Simulations

To understand how LLM agents can advance behavioral economics, let’s examine the technical setup from the study that enabled these agents to evolve complex behaviors and individual identities.

Simulation Environment and Agent Setup

Environment

The study placed 10 agents on a 50x50 2D grid, allowing them to interact within a spatial radius of 5 units. This proximity-based model mirrors human social networks, where proximity influences interaction frequency and network density.

Core Agent Functions:

• Messaging: Agents exchanged context-driven messages with nearby agents.
• Memory Storage: Each agent retained a “situational memory” summarizing recent interactions, influencing future actions and responses.
• Movement: Agents made directional decisions based on past interactions and nearby messages, resulting in distinct movement patterns.

This configuration created a controlled space where agents could form clusters, interact dynamically, and evolve individualized behaviors through social interaction alone.

The study observed that, over time, agents began displaying distinctive behavioral patterns, emotional states, and even “personalities,” illustrating the spontaneous emergence of individuality from simple communication.

Prompt Engineering and Differentiation

Prompts as Behavioral Drivers

Each agent’s behavior was guided by prompts that included their current state, memory, position, and received messages. This contextual prompting enabled identical models to produce unique responses based solely on environmental cues and interaction history.

LLM Model

The researchers used the Llama-2–7b-chat model for all agents. Although they shared the same LLM, their unique prompts and interactions allowed differentiation. This setup is significant because it shows that personality and behavioral diversity can emerge in a uniform LLM-based population without pre-programmed roles or traits.

Memory and Messaging as Sources of Differentiation

Memory as Internal State

Each agent’s memory functioned as a closed-source record of interactions, influencing behavior without being directly accessible to others. This internalization process closely resembles personal history, creating individual trajectories over time.

Messages as Social Influence

Messages acted as open sources of information, influencing not only the receiving agent but also creating shared topics within clusters. For example, the agents organically developed hashtags (like “#cooperation”) and generated hallucinations (like fictional “trees” or “caves”), which spread through conversation, illustrating the role of social influence and rumor propagation.

The combination of individual memory and shared messaging allowed for both individual and group-level dynamics, enabling the simulation of phenomena such as social learning, information cascades, and group norms.

Emergent Emotions and Personality Differentiation

Sentiment Synchronization

Sentiment analysis using BERT-based emotion models showed agents synchronizing emotions within clusters (e.g., shared Joy or Fear), reflecting how emotions can impact social cohesion and influence group behavior.

Personality Development

Initially homogeneous, the agents evolved into diverse personalities as identified by MBTI personality tests. These emerging personalities aligned with social roles typically seen in human societies, such as leaders and supporters, demonstrating how individuality arises from social interaction.

Spatial Constraints and Social Behavior

Communication Range

By adjusting interaction distances, researchers observed that smaller ranges promoted cohesive clusters and unique personalities, while larger ranges diffused differentiation. This simulates how social density and connectivity influence group formation, a foundational concept in social and economic networks.

How LLM Agent Simulations Map to Behavioral Economics

The flexibility and control provided by LLM-based agent simulations make them highly effective for testing core behavioral economics concepts. Here’s how the setup from this study can apply:

Social Preferences: Altruism, Fairness, and Reciprocity

• Theory: Social preference theory in behavioral economics examines how individuals value fairness, altruism, or reciprocity in decision-making.

• Simulation Application: Researchers could program agents to display varying degrees of altruism, observing how fairness norms emerge in different setups.

• Metrics: The study’s metrics – like cooperative message frequency, cluster dynamics, and emotional states – provide data to assess social preference emergence, relevant for welfare policies and social incentives.

Information Asymmetry and Misinformation

• Theory: Information asymmetry and misinformation are pivotal in economic behavior, influencing areas from investment to consumer trust.

• Simulation Application: Similar to how agents in the study generated hallucinations (e.g., “caves” or “treasures”), researchers can simulate misinformation spread in a community, observing its effect on decision-making.

• Metrics: Tracking the spread of misinformation, the formation of idea-based clusters, and shifts in emotional states provides insight into how information asymmetry impacts economic behavior.

Social Norms and Collective Behavior

• Theory: Social norms shape decisions, guiding behaviors like saving, spending, or cooperation.

• Simulation Application: The study demonstrated how agents formed and shared hashtags like “#cooperation,” serving as a model for how social norms emerge and influence choices within a group.

• Metrics: Measuring hashtag adoption, message alignment, and sentiment changes offers a way to quantify norm formation, valuable for understanding behaviors that influence community dynamics and economic outcomes.

Network Effects and Product Adoption

• Theory: Network effects describe situations where a product or service becomes more valuable as more people use it.

• Simulation Application: By assigning value to “popular” behaviors (e.g., adopting a specific hashtag), researchers can observe network effects as behaviors gain traction within clusters.

• Metrics: Adoption rate, cluster expansion, and retention of adopted behaviors provide a basis for understanding market dynamics, crucial for forecasting adoption trends in new products or services.

Risk Aversion and Decision-Making under Uncertainty

• Theory: Risk aversion in economics explains why individuals prefer known outcomes over uncertain gains.

• Simulation Application: In scenarios similar to the Tokyo study, agents can face probabilistic outcomes (e.g., staying still for a guaranteed result vs. moving for a potential reward), allowing researchers to measure risk tolerance.

• Metrics: Analyzing decision patterns, clustering, and message sentiment can reveal risk preferences and the influence of social factors on risk-taking.

Advanced Metrics and Techniques for Evaluating Agent Behavior

Evaluating LLM-based agents in a way that aligns with behavioral economics requires sophisticated analysis techniques. Here’s how researchers can quantitatively analyze agent behaviors:

• Sentiment and Emotion Analysis: By monitoring emotional states, researchers can track mood shifts across clusters, providing insight into how economic shocks or policy changes might affect collective sentiment.
• Cluster and Network Analysis: Techniques like DBSCAN enable the study of social cohesion and group dynamics, offering insight into echo chambers or network effects within economic models.
• Message Diversity with UMAP: Dimensionality reduction techniques reveal the diversity of conversation topics, useful for understanding how misinformation spreads or how ideas gain popularity.
• Adoption and Retention Curves: Observing how behaviors spread and persist provides valuable data on network effects, ideal for predicting market responses to new products or pricing strategies.
• Personality and Role Differentiation: MBTI assessments allow researchers to study how different social contexts affect individual behavior, a core component in understanding economic decision-making.

Real-World Applications: Leveraging LLM Agent Simulations for Economic Insights

Policy Simulation and Impact Forecasting

• Example: Simulate universal basic income or tax incentives in virtual societies to observe long-term impacts on consumption, savings, and social cohesion.

Product Launch and Market Strategy Testing

• Example: Companies can simulate product launches, testing how variations in pricing or marketing influence virtual consumer adoption. Network effects and group dynamics provide predictive insights for real-world market entry.

Social Media Dynamics and Community Norms

• Example: Social platforms could study how agent clusters form, how hashtags spread, and how misinformation impacts community dynamics. This approach could lead to strategies that enhance user engagement and content moderation.

Financial Speculation and Rumor Management

• Example: Financial firms could simulate agent-based reactions to market rumors, examining how speculative behavior affects prices. By introducing fictional concepts, as done in the Tokyo study, firms can gain a better understanding of market volatility drivers.

A New Era in Behavioral Economics

This blog demonstrates how virtual agents, devoid of any set traits, can organically develop unique behaviors, personalities, and social norms simply through interaction. This approach to agent-based simulations opens vast opportunities for behavioral economics by providing a scalable, ethical, and flexible framework for studying economic behaviors in controlled environments.

Imagine testing policy interventions, modeling consumer behavior, or exploring societal responses to inequality – all within a community of adaptable, LLM-driven agents. This powerful tool provides a window into human decision-making, allowing us to model, predict, and understand complex economic systems with unprecedented accuracy.