Spontaneous Emergence of Agent Individuality through Social Interactions in LLM-Based Communities (2411.03252v1)

Abstract: We study the emergence of agency from scratch by using LLM-based agents. In previous studies of LLM-based agents, each agent's characteristics, including personality and memory, have traditionally been predefined. We focused on how individuality, such as behavior, personality, and memory, can be differentiated from an undifferentiated state. The present LLM agents engage in cooperative communication within a group simulation, exchanging context-based messages in natural language. By analyzing this multi-agent simulation, we report valuable new insights into how social norms, cooperation, and personality traits can emerge spontaneously. This paper demonstrates that autonomously interacting LLM-powered agents generate hallucinations and hashtags to sustain communication, which, in turn, increases the diversity of words within their interactions. Each agent's emotions shift through communication, and as they form communities, the personalities of the agents emerge and evolve accordingly. This computational modeling approach and its findings will provide a new method for analyzing collective artificial intelligence.

Emergence of Agent Individuality in LLM-Based Simulations

The paper "Spontaneous Emergence of Agent Individuality through Social Interactions in LLM-Based Communities" explores how individuality emerges in homogeneous LLM-based agents through social interactions. This paper sheds light on the differentiation of behaviors, emotions, and personalities without any predefined characteristics or initial memories within the agent community. The findings add to the understanding of collective intelligence and the dynamics of agent interactions.

The primary focus of the paper is to investigate how individual traits and collective behaviors arise spontaneously when multiple LLM agents interact in a shared environment. To simulate these dynamics, the researchers deployed 10 LLM agents in a 2D environment. These agents were equipped to send messages and receive information from nearby agents, store situational summaries, and make movement decisions. The simulation relied on the Llama 2 model, an open-source LLM, as the underlying linguistic engine.

A key result from this paper is the noticeable differentiation in behaviors exhibited by the LLM agents. The agents began with no distinct individual properties, yet over time, distinct personalities emerged as they communicated and interacted. Notably, agents exhibited recursive habits such as clustering and generated unique "stay" commands based on their interaction histories. The simulation also demonstrated a bias in movement commands, suggesting that internal architectures and training data might influence emergent behavior patterns.

The paper also highlights that while the agent's memories remained distributed and diverse, their messages demonstrated thematic convergence within clusters. Using Sentence-BERT analysis, the research shows that messages became more uniform as agents grouped together, reflecting an emergent coordination in language and thought, despite each starting from a similar initial state.

Moreover, distinct emotional states were tracked within the agents, indicating that their emotional expressions, derived from natural language processing, were synchronized within clusters but varied across different groups. This indicates that emotional states can become aligned through social interactions, similar to biological entities.

The paper further observes the phenomenon of hallucinations, a term referring to the creation of content not present in the prompts. For example, words such as “hill” or “cave system” emerged, indicating that these agents were capable of creating imaginative constructs during communication. In addition, the propagation of hashtags suggests the development of shared narrative elements within agent groups, emphasizing the formation of nascent social norms.

Personality differentiation among agents was quantitatively assessed using the Myers-Briggs Type Indicator (MBTI) test. Although initial tests reflected similar personality types across agents, subsequent interactions led to a diversification of types, reflecting broad personality shifts and the natural adoption of various roles within the community.

Spatial scale played a crucial role in shaping the dynamics of agent behavior. Increased spatial scale correlated with heightened diversity in message content and emergence of hallucinations, whereas unique hashtag formation diminished. This suggests a trade-off between broad communication and the maintenance of shared themes.

In conclusion, the paper effectively demonstrates that individuality and collective behavior can naturally emerge in LLM-based agent communities through interaction. These findings have implications for the paper of artificial societies and collective intelligence. They provide a model for examining how autonomous agents might develop complex behavioral patterns without explicit coding of individual traits or social rules. Future work may build on these findings to explore broader implications in the field of artificial intelligence, specifically in the domains of automated decision-making systems, interactive AI agents, and social robotics.