arXiv cs.AITuesday · May 26, 2026FREE

In Search of the Ingredients of Open-Endedness: Replicating Picbreeder with Large Vision-Language Models

A new arXiv paper (2605.23908) from researchers attempts to replicate Picbreeder, a classic human-driven open-ended search system, using frontier vision-language models (VLMs) instead of human users. Picbreeder originally allowed users to collaboratively evolve images through interactive evolution of small neural networks, generating a diverse library. The study replaces human selectors with VLMs and observes clear qualitative differences in the output compared to the historical human baseline. To characterize these differences, the authors employ metrics of phylogenetic complexity, visual salience, and semantic novelty. They also investigate the effect of adding exploratory noise to the agents' selection process as a potential causal factor. The work aims to understand whether artificial agents can exhibit the open-endedness characteristic of human creative processes. The paper was published on arXiv on May 26, 2026, under the cs.AI category.