In Search of the Ingredients of Open-Endedness: Replicating Picbreeder with Large Vision-Language Models
The paper explores the potential of large Vision-Language Models (VLMs) to replicate the open-ended creative processes exemplified by Picbreeder. It highlights the differences in output between AI-driven and human-driven creative processes, focusing on metrics such as complexity and novelty. The authors investigate factors that may influence these differences, including exploratory noise and behavioral diversity among agents.
- ▪The study aims to understand if artificial agents can achieve open-ended discovery similar to humans.
- ▪Picbreeder serves as a model for human-driven open-ended search, generating diverse images through collaborative evolution.
- ▪The research identifies qualitative differences in outputs between VLMs and human users, analyzing them through various metrics.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.23908 (cs) [Submitted on 1 Apr 2026] Title:In Search of the Ingredients of Open-Endedness: Replicating Picbreeder with Large Vision-Language Models Authors:Sam Earle, Kay Arulkumaran, Andrew Dai, Akarsh Kumar, Julian Togelius, Sebastian Risi View a PDF of the paper titled In Search of the Ingredients of Open-Endedness: Replicating Picbreeder with Large Vision-Language Models, by Sam Earle and 5 other authors View PDF HTML (experimental) Abstract:We are in the midst of large-scale industrial and academic efforts to automate the processes of scientific, technological and creative production through AI-driven assistants.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
