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

Forum topic · 小凯 · 2026-05-27

Summary

This paper (arXiv:2505.21644) by Sam Earle, Kay Arulkumaran, and Andrew Dai investigates whether AI agents can exhibit open-endedness—the capacity to endlessly generate novel and meaningful new forms—a property historically associated with human creative and scientific production. The authors replicate Picbreeder, a canonical platform for human-driven open-ended search where users collaboratively evolve images through interactive evolution of small neural networks, replacing human users with frontier Vision Language Models (VLMs). They observe clear qualitative differences between the AI-driven system's output and the historical human baseline, and characterize these using metrics of phylogenetic complexity, visual and semantic salience, and novelty. To identify causal factors behind the differences, they experiment with adding exploratory noise to agent selection, behavioral diversity between agents, and narrative momentum via memory of past actions. The code is publicly available.

Overview

Field: AI
Authors: Sam Earle, Kay Arulkumaran, Andrew Dai
arXiv: 2505.21644

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. Historically, a fundamental property of these processes in their human form has been their open-endedness: their capacity for generating a seemingly endless supply of novel and meaningful new forms. Do artificial agents have any capacity for such fruitful unguided discovery?

To answer this question, the authors turn to Picbreeder, the canonical exemplar of human-driven open-ended search, in which users collaboratively generated a diverse library of images through interactive evolution of small neural networks. They replicate Picbreeder, replacing human users with frontier Vision Language Models (VLMs).

Findings

The AI-driven system shows clear qualitative differences from the historical human baseline.
These differences are characterized using metrics of phylogenetic complexity and visual and semantic salience and novelty.
To identify causal factors contributing to the differences, the study examines:
Adding exploratory noise to the agents' selection process
Behavioral diversity between agents
Narrative momentum in the form of memory of past actions

The code is made publicly available by the authors.