Picbreeder replayed with VLMs to probe AI open-endedness

A group of researchers took Picbreeder, the interactive-evolution site where humans once co-authored a library of images by picking parents from slates of small neural-network offspring, and replaced the humans with frontier vision-language models. Their arXiv paper, by Sam Earle, Kai Arulkumaran, Andrew Dai, Akarsh Kumar, Julian Togelius and Sebastian Risi, describes a working replica of the platform with VLMs sitting where the users used to sit.

The machine-driven run does not look like the human one. The authors report "clear qualitative differences between the output of our system and the historical human baseline", and try to pin those differences down with metrics of phylogenetic complexity and visual and semantic salience and novelty.

From there they probe what might be missing. The paper studies "the addition of exploratory noise to the agents' selection process, of behavioral diversity between agents, and of narrative momentum in the form of memory of past actions." The framing is diagnostic: three candidate ingredients of open-endedness, tested one at a time against the historical human trace.

The abstract does not publish per-metric numbers or say which VLMs stood in for the crowd. Code is on GitHub, and the paper is bound for GECCO 2026 in San José, Costa Rica. Two experts in our Who's Who directory have shared it.