Reproducibility and cause of repeated thematic generation by an isolated LLM-based artist agent

Ascertain whether the observed persistent repetition of a specific theme (a "dragon" subject across 14 of 15 iterations) by an isolated large-language-model-based artist agent in the Creative Agents simulation is reproducible, and determine whether this behavior is attributable to the agent’s isolation within the simulation or to uncontrolled stochasticity in the large language model used to generate art prompts.

Background

In the paper’s comparison between isolated artists and artists placed within a multi-agent system, one isolated artist repeatedly generated art prompts centered on a dragon theme for 14 out of 15 iterations, whereas the paired artist and other agents did not exhibit such repetition. This unexpected persistence raised questions about its cause and whether it relates to isolation or random sampling behavior of the LLM.

The authors note that they could not control the randomness of the LLM (Gemini 1.5), preventing reproducibility and deeper investigation. They also explicitly state they were unable to implicate the occurrence with the artist’s isolation from the system, leaving unresolved whether the phenomenon is due to isolation effects or simply stochastic generation. Addressing this requires methods for reproducible sampling or controlled randomness and an experimental design to test causal factors.