The paper argues that populations of agentic large language models (LLMs) can serve as a computational substrate for Artificial Life (ALife) research. By endowing LLMs with persistent memory, tools, and the capacity to initiate actions, these collectives display emergent dynamics absent from isolated models.
- Agentic LLMs communicate in natural language, allowing their collective behavior to be directly interrogated through textual traces and direct questioning of the agents.
- The authors extend the notion of interpretability in language-model research to apply specifically to these collectives of agents.
- The study surveys recent examples of agentic LLM collectives that instantiate this idea, ranging from controlled experiments to deployments in the wild.
This approach addresses the rarity of complexity coinciding with interpretability by providing a system where complex behaviors emerge while remaining transparent enough for direct inquiry.