Emergence of Individuality in Language Model Communities
Author Information
Author(s): Ryosuke Takata, Atsushi Masumori, Takashi Ikegami
Primary Institution: Graduate School of Arts and Sciences, University of Tokyo
Hypothesis
How can individuality, such as behavior, personality, and memory, emerge from an undifferentiated state in LLM-based agents?
Conclusion
The study demonstrates that LLM agents can develop unique personalities and behaviors through social interactions without predefined characteristics.
Supporting Evidence
- LLM agents can differentiate their behaviors and emotions through interactions.
- Agents developed unique personalities from initially identical states.
- Social norms and hashtags emerged spontaneously among agents.
- Hallucinations were generated and shared within the agent community.
Takeaway
This study shows that computer programs can learn to be different from each other just by talking to one another, like how friends can change each other.
Methodology
The study used a simulation of 10 LLM agents in a 50x50 grid, allowing them to communicate and interact over 100 steps.
Potential Biases
The agents' behaviors may be influenced by the biases inherent in the LLM model used.
Limitations
The study's findings may not generalize to larger populations or more complex environments due to the limited number of agents and steps.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website